neural-bridge.dev

// Live feed

Build log

Chronological feed of work across my projects. Milestones, releases, hardening, posts. Hand-curated for now; auto-sync from merged PRs lands later.

November 2026 · 8 entries

  1. // note · latest neural-bridge-blog

    Connect Project Husky page to Husky Lab Mac mini deployment

    • Sets siteUrl: "https://husky.neural-bridge.dev" in src/content/projects/project-husky.md and src/content/projects/ko/project-husky.md so the Project Husky header includes a SITE → link alongside a dedicated ## Current access (## 현재 접속 방식) section modeled on moonbase.md. - Adds **[Launch Husky…
  2. // feature neural-bridge

    Md_to_pdf.py helper (weasyprint, CJK-safe)

    Adds scripts/md_to_pdf.py: a one-command Markdown-to-PDF helper.

  3. // note neural-bridge

    /discuss back-and-forth, @-tag routing, typing indicator

    Three problems with the Herman council bot, all seen on 2026-11-02:

  4. // note neural-bridge

    Add Luna as the executive-assistant lane, and a room log

    Andy wants Luna, his executive assistant, in the Herman council alongside Yor and Loid, and wants everything said in the room to reach Yor’s nightly memory, including turns where Yor didn’t speak.

  5. // note neural-bridge

    Yor runs from a neutral folder with her DM toolsets

    • Hermes loads project context files (AGENTS.md and the like) from the working directory. The council spawned Yor’s hermes chat -q one-shot from this repo, so its schema joined her identity prompt: 51,215 characters against 29,175 from a neutral folder (measured 2026-11-02). - The CLI’s default toolset bundle…
  6. // note neural-bridge

    An advisor Andy names answers instead of passing

    On 2026-11-02 Andy sent a short hello tagged to Luna. She was invoked and replied [PASS], which the council posts as nothing, so it looked as if she had not been added to the room.

  7. // note neural-bridge

    Daily meeting notes in the vault, with a nightly summary

    The council now keeps a daily meeting note in the vault (one note per day in the Council folder), rebuilt from the room log after every exchange and every /discuss, plus a short end-of-day summary written by a small model at 23:30.

  8. // note neural-bridge

    Shared files and Google Drive links the advisors can read

    Andy can now send a PDF, Word, PowerPoint or text file to the council room, or paste a Google Docs, Sheets, Slides or Drive link, and Yor, Loid and Luna can read it for three hours.

October 2026 · 29 entries

  1. // note neural-bridge

    Make "it's done" stick on the board

    Andy told Luna three times that a set of overdue cards were done. She said “logged” each time and made no tool call. The cards stayed open on Herman Tasks.md, which is what her check-ins, the Needs Andy digest and the Gate Registry all read, so the check-ins kept raising them for a week.

  2. // note neural-bridge

    One-shot calls: drop Claude Code's tool and skill scaffolding

    Cost fix #7 in reports/Neural Bridge fleet API costs.md. Every claude -p call loads Claude Code’s built-in tool definitions and skill listing, even when it never uses a tool.

  3. // note neural-bridge

    Flush only what a session added since its last flush

    Cost fix #6 in reports/Neural Bridge fleet API costs.md (the delta half of lever 3). A resumed conversation (Luna on Telegram, every Discord agent) keeps one session id, and Claude Code appends each turn to the same JSONL file. Every turn’s flush sent that whole file, capped to its last 400K characters: raw…

  4. // note neural-bridge

    Stop paying twice for failures

    Step 5 in reports/Neural Bridge fleet API costs.md. Two paths bought a second full-price call after any failure.

  5. // note neural-bridge

    Log every fleet claude -p call: job, route, tokens, cost

    Step 4 in reports/Neural Bridge fleet API costs.md. Since the fleet moved to the API key (#208), every cost figure has been an estimate rebuilt from transcripts at list prices. Transcripts can’t say which job a session was or which route paid for it. The caller knows both, and Claude Code reports its own usage and…

  6. // note neural-bridge

    Resumed sessions send only what changed, and rotate past 100K tokens

    Step 11 in reports/Neural Bridge fleet API costs.md. A resumed session keeps every prompt it was ever sent, and every agent turn re-sent the whole mention prompt. For Luna that’s about 81,000 characters, 75,000 of them identical every turn: the charter (42K), notes (12.6K), the template’s rules (13.6K), the voice…

  7. // note neural-bridge

    Usage log: one call's cost per line, not the session's running total

    Found while verifying the #217 deploy. For a resumed session, Claude Code’s JSON result reports the session’s running totals, not the call’s.

  8. // note neural-bridge

    Fleet cost quick wins: no unread flushes, no doubled prompts, capped agentic runs

    The first three fixes from the fleet cost audit (reports/Neural Bridge fleet API costs.md, 2026-10-17). Since #208 the fleet bills Andy’s own Anthropic API key, estimated at about $232 a month at list prices.

  9. // note neural-bridge

    Loop engineer: tell Andy through the review queue, not Discord

    • No more Discord posts from the loop. Gone: the post on every claim, PR, block and escalation, the daily-cap notice, and the end-of-run summary that went out every hour, even when nothing had happened. - An issue the loop stops on becomes a review queue alert card. - It covers agent-blocked (for example,…
  10. // note neural-bridge

    Run the fleet on the Anthropic API instead of the Copilot proxy

    On 2026-10-16 GitHub began refusing the Copilot token exchange behind copilot-api, and Copilot then rejected every Claude 5 request (model_not_supported). Fleet turns on Claude 5 fell back to Opus 4.8 or failed, and the gateway’s Opus 4.8 requests started returning 400 too.

  11. // note neural-bridge

    Promote fresh-session-stale-state-restatement, generalized for publishing

    The nightly compile promoted this concept on 2026-10-14 (3-vote PROMOTE, from a Luna session) and left it uncommitted in the shared checkout, which blocked auto-reload from deploying #207 until it was stashed.

  12. // note neural-bridge

    Nightly compile: stage in private, file through the review queue

    • New scripts/review_queue/promotions.py, the 03:00 job’s new entry point. - compile.py runs unchanged on a private copy of knowledge/ at .trees/compile-staging, through the new NB_COMPILE_KNOWLEDGE_DIR. - Every article it promotes or quarantines becomes a capture card: - full text in the local queue…
  13. // note neural-bridge-blog

    Publish Moonbase Mac mini AI Command Deck articles and project hub update

    Publishes the two new Moonbase companion articles and updates the Moonbase project hub on neural-bridge.dev following the Moonbase Mac mini AI Command Deck release (andy-herman/moonbase#20 and andy-herman/moonbase#22):

  14. // note neural-bridge-blog

    Gemma GRC spec v1.3: Gemma 4 E4B adopted after Runs 013 and 013b

    Updates the unlisted Gemma GRC technical specification to v1.3. It stays draft: true and publishes on 2026-10-19 through the Tuesday queue.

  15. // note neural-bridge-blog

    Gemma GRC project page: Gemma 4 E4B adopted

    Brings the Gemma GRC project page (English and Korean) up to date with the generator swap. This page is live, so merging publishes it. Unlike the spec, it is not a draft.

  16. // note neural-bridge-blog

    Publish Moonbase persistent workbench and technical specification

    Publish the Moonbase Workbench build article and standalone v1.0 engineering specification dated October 13, and update the existing Moonbase project hub without changing historical URLs.

  17. // note neural-bridge

    Review queue, slice C: the loop engineer's PRs, merged from Telegram once their checks pass

    Slice C of the Phase 1 review queue. Andy’s decision on 2026-10-12: approving a loop-engineer PR merges it, but only once its checks are green.

  18. // note neural-bridge

    Review queue, slice D part 1: this week's post as a card, and the queue in Luna's morning check-in

    (scripts/review_queue/publish.py) - Producer: prepare_week.py raises one publish card for the post the Monday run will publish, keyed by slug and week. It is idempotent and never fails the prep. Title: “Publishes Mon 19 Oct: ”. The card links to the draft on GitHub and points to the LinkedIn and X…

  19. // note neural-bridge-blog

    Gemma GRC spec v1.2: extended eval and the prompt experiments

    Updates the queued research draft gemma-grc-technical-specification to v1.2. It is still draft: true for 2026-10-19.

  20. // note neural-bridge

    Review queue, slice B: questions answered from Telegram, and PR proposals that survive a restart

    Slice B of the Phase 1 review queue: questions and actions.

  21. // note neural-bridge

    Flush no longer dies on long sessions, and questions from a live chat wait until it goes quiet

    Found while testing slice B live on 2026-10-12.

  22. // note neural-bridge

    Progress log: leave questions answered on a card out of the Open list

    • questions.unanswered(agent, session_id, open_questions) returns a session’s open questions minus those Andy answered or dismissed on a card in that session. It uses the same word-overlap match (same_question, 0.6 or more) that already stops such a question from being raised as a new card. - hooks/flush.py…
  23. // note neural-bridge

    Review queue cutover: alerts go to Luna's bot, and the canary judges each store on its own readers

    • Alerts are pushed. health.DEFAULT_PUSH_KINDS is now (capture, alert), so Luna’s bot sends alert cards with Acknowledge and Snooze, the same way it sends captures. - The canary stops crying wolf about Luna. luna_notes, luna_live_state and echo_voice now name their readers in WATCHED. Each stays…
  24. // note neural-bridge-blog

    Gemma GRC spec v1.1: Run 011 results

    Updates the queued research draft gemma-grc-technical-specification to v1.1. It is still draft: true for 2026-10-19.

  25. // note neural-bridge

    Add a finite source-only Moonbase stdio work bridge

    • Add fixed python -m scripts.moonbase_bridge catalog|run entry points using the released plugin/Discord registry, without changing the existing telemetry producer or schema. - Implement project-check for automation-engineer as an audited Python filesystem scanner, not model work or a project-command runner.…
  26. // note neural-bridge-blog

    Make agent profiles job-first

    • Present agent roles and useful jobs before portraits and biography, using the reviewed English/Korean copy for Luna, Loid and Professor. - Add strict, localized five-detail job cards and compatible optional account/model/source metadata. Hide unqualified model claims and use an intentional gray initial for Loid. -…
  27. // note neural-bridge-blog

    Husky project page: say "the full test suite", not "394 tests"

    andy-herman/project-husky#1 and #2 are merged into main, so CI now runs on every pull request there. The suite has outgrown 394 tests: Windows ran 476 on #2, and #3 brings it to 626.

  28. // note neural-bridge

    Proxy watchdog: catch a stale model connection before an agent does

    The model gateway now checks its upstream connection every five minutes. It restarts the connection before its token expires, and it catches refusals from any tool that uses it, not only the fleet. It raises a review-queue alert when a restart doesn’t help. Design and runbook: the new “proxy watchdog” section in…

  29. // note neural-bridge

    Publish prep on Sonnet 5, with a fallback

    The Sunday publish prep now drafts each week’s LinkedIn post on Claude Sonnet 5, the same model as the content agent. If that fails, it falls back to Opus 4.8, so the prep never comes back empty.

September 2026 · 40 entries

  1. // note neural-bridge-blog

    Publish "The Model Is Supposed to Fail" and add Project Husky

    • Post: src/content/posts/the-model-is-supposed-to-fail.mdx, “The Model Is Supposed to Fail”. About 2,400 words, five runtime captures, and a banner. - Publishing on 2026-09-30 at Andy’s request, instead of queueing for 2026-10-26. - The file keeps draft: true with pubDate: 2026-09-30. After merge, a manual…
  2. // note neural-bridge-blog

    Link private Moonbase workspace and settings

    Update only the existing Moonbase project’s Current access section at https://neural-bridge.dev/projects/moonbase/:

  3. // note neural-bridge

    Report the real cause of a failed claude -p turn, not the connectors warning

    When a Luna turn failed on 2026-09-30, the error the daemon posted to Discord named a harmless warning and cut off the real cause. This makes both claude -p callers, the daemon and the flush hook, report the line that carries the API error.

  4. // note neural-bridge

    Model gateway: Claude 5 for the fleet, a self-healing model connection, no stray MCP tools

    A local model gateway that gives the Neural Bridge fleet the Claude 5 models, repairs its model connection on its own when it goes stale, and stops every agent turn from loading about 150 unrelated tools. Design and runbook: docs/MODEL_GATEWAY.md.

  5. // note neural-bridge

    Luna and research on Claude Opus 5, falling back to Opus 4.8 on failure

    Luna and the research agent now run on Claude Opus 5. Every other agent keeps the current default, Opus 4.8. If an Opus 5 turn fails, it is answered on Opus 4.8 instead of failing.

  6. // note neural-bridge

    Writing agents on Claude Sonnet 5

    The writing agents (content, social and docs-editor) now run on Claude Sonnet 5. Luna and research stay on Opus 5 (#192), and every other agent stays on the fleet default, Opus 4.8.

  7. // note neural-bridge-blog

    Correct the Husky post and project page on student data

    The FERPA review of Project Husky (andy-herman/project-husky#2) checked the live post and project page against the code. Five statements needed fixing. This PR corrects them in English and Korean.

  8. // release neural-bridge

    The review queue is live: agent work now comes to me

    Phase 1 of the August roadmap, first slice. The agents worked; what they lacked was a trigger. Every one of them waited for me to start the conversation, and the Discord fleet sat unused for weeks.

    Now agent output that needs me lands in one local queue. Luna’s Telegram bot sends each item as a card with buttons, and every decision and its outcome goes into an append-only event log. Two kinds of item are live:

    • Captures, from the filing gate’s quarantine. Filing one opens a pull request that moves it into the wiki and merges it. Rejecting one removes it. Both take a second tap.
    • Alerts, from the memory canary, the deploy watcher and the scheduled-output watchdogs. For the first week they are collected but not sent as cards, so nothing arrives twice.

    On its first night the queue sent eight capture cards, for concepts that had sat in quarantine since 2026-05-10. Nothing had ever asked anyone to decide them.

    The queue checks itself. Every day a probe item goes through the whole lifecycle, and the pusher’s heartbeat must be fresh. A separate watcher alerts directly if the queue stops delivering, because a queue that has stopped cannot report its own failure. The health check had a bug of its own, found during deployment: two checks in the same second shared a probe key, and the second reported a healthy queue as broken. PR #176 fixed it.

    The measure after four weeks: decisions per week, median time to decision and how many items expire unread.

  9. // note neural-bridge-blog

    Post banners: optional banner image and social preview

    Adds an optional banner to posts.

  10. // note neural-bridge-blog

    Bring the site up to date: Neural Bridge page, build log, About, Luna

    Brings neural-bridge.dev current as of 2026-09-29.

  11. // note neural-bridge-blog

    Queue "Gemma GRC: A Technical Specification" for 2026-10-19

    Adds a long-form working paper, src/content/research/gemma-grc-technical-specification.mdx, queued as a draft for Monday 2026-10-19 (the next free slot after 10-05 and 10-12).

  12. // note neural-bridge-blog

    Publish Moonbase and unlist the cafe project

    • Publish Building Moonbase: a small world for real agents dated 2026-09-29 and add the Moonbase project page. - Include three genuine, full-frame application captures, a responsive habitat banner, and three Mermaid data-flow diagrams rendered as static light/dark SVGs. - Preserve Agent Office’s pinned MIT…
  13. // note neural-bridge

    Promote ai-enabled-social-engineering-canon from quarantine (review queue ec3f45)

    Decided by Andy in the review queue (item ec3f45).

  14. // note neural-bridge

    Promote cloud-native-firewalls-replace-appliances from quarantine (review queue 5608e0)

    Decided by Andy in the review queue (item 5608e0).

  15. // note neural-bridge

    Promote docker-up-doesnt-return from quarantine (review queue 66429b)

    Decided by Andy in the review queue (item 66429b).

  16. // note neural-bridge

    Promote defenders-can-win-stories from quarantine (review queue 283512)

    Decided by Andy in the review queue (item 283512).

  17. // note neural-bridge

    Promote hashing-vs-encryption-explicit-distinction from quarantine (review queue 1632e5)

    Decided by Andy in the review queue (item 1632e5).

  18. // note neural-bridge

    Promote idor-still-a01-in-2025 from quarantine (review queue b13d83)

    Decided by Andy in the review queue (item b13d83).

  19. // note neural-bridge

    Promote lockout-vs-credential-stuffing from quarantine (review queue 74d6cc)

    Decided by Andy in the review queue (item 74d6cc).

  20. // note neural-bridge

    Promote vdp-vs-bug-bounty-distinction from quarantine (review queue ef3a32)

    Decided by Andy in the review queue (item ef3a32).

  21. // note neural-bridge

    Add opt-in local specialist telemetry for Moonbase

    • Add an opt-in, private, atomically replaced JSON snapshot with a dynamically discovered specialist registry and independent Discord/Telegram writer slots. - Observe model-subprocess attempts, overlapping work, cancellation, connection state, and lost observations without changing tools, permissions, approvals,…
  22. // note neural-bridge

    Support persistent telemetry opt-in via allowlisted env loading

    Source-only follow-up to #185.

  23. // note neural-bridge

    Honcho kept on probation: gate G3 measures freshness, peer card sanitized

    Consolidation Step 4 (keep or demote Honcho) re-examined. The decision is keep injected, on probation, with two instruments the gate lacked, and a dated record in docs/MEMORY_CONSOLIDATION.md of what decides it and when.

  24. // note neural-bridge

    Fix binary Git patch decoding in the outbound guard

    • Capture Git stdout as bytes and decode UTF-8 with replacement instead of failing on binary bytes. - Parse patches at Git’s LF boundaries, preserving valid text after embedded control characters and Unicode separators. - Keep the same scanner, forced --text, all unpushed history, marking/shingle policy, and…
  25. // note neural-bridge

    Screen the loop engineer's push with the outbound guard

    The loop engineer (scripts/loop_engineer/) pushed its branch and opened a PR with no screening. It was the last route from agent-written text to GitHub outside the outbound guard (#167). The branch and the PR are public the moment they exist, the PR body carries the agent’s own summary, and the agent may have read…

  26. // note neural-bridge

    SessionStart injects progress.md instead of raw daily-log files

    Retires the SessionStart daily-log injection in favour of progress.md, the candidate named in STATUS.md’s known gaps. This completes Step 1 of docs/MEMORY_CONSOLIDATION.md literally: both vault artifacts (notes.md, progress.md) are now re-read at session start.

  27. // note neural-bridge

    Add a git pre-push hook as the outbound guard's catch-all

    The outbound guard ran only where code called it: compile, the daemon’s GitHub actions and the loop engineer. A push made any other way went out unscreened. That includes a manual push of text an agent wrote, and a push from a script that forgot to call the guard.

  28. // fix neural-bridge-blog

    Keep draft posts and research out of search indexes

    Draft posts and research build and deploy at their final URLs until the Monday publish cron flips them. The listing pages and RSS feed already skip drafts, but search engines could still find them:

  29. // note neural-bridge-blog

    Projects dropdown with Gemma GRC, Bellwether and a 1920s café in Seoul

    Puts the key projects in the nav. Projects becomes item 02, a dropdown of featured projects with one-line taglines, and each project page gets a data-flow diagram.

  30. // note neural-bridge-blog

    Correct what ran on the device in the Gemma GRC paper

    The published paper said all training, generation, judging and inference ran on the device, then described a cloud teacher model writing the first day’s question-answer pairs. This makes the intro and abstract consistent with that, says local generation has kept chunks on the machine since, and aligns the…

  31. // note neural-bridge

    Owner-only private MCP tools for agent turns

    Some MCP tools reach private material on the owner’s Mac. They must never be reachable by: - readers of this public repo; - agent-to-agent handoffs; - other processes that happen to run claude.

  32. // note neural-bridge

    Fix the memory pipeline's write path and finish consolidation Step 4

    Deploying #162/#163 on the Mac (after #164 revived auto-reload) showed flush_daily_log FAILING. The write side of the memory pipeline has never worked for agent turns:

  33. // note neural-bridge

    Outbound guard for marked vault text on the wiki and GitHub routes

    Agents read the Obsidian vault, and two routes publish what they write: the wiki (knowledge/, tracked in this repo) and GitHub (issues, comments, PR branches and bodies). The filing gate screens for poisoning, not confidentiality. Nothing screened outbound text for material from marked notes. A read-only leak check…

  34. // note neural-bridge

    Exempt already-published text from the outbound guard index

    Widening the private outbound-guard policy made 4 of the 219 published blog files match the index. Three matched on public NIST citation lines (the AI 600-1 PDF, the AI RMF URL) that a marked note also holds, and one on a published 14-word phrase. Every later post citing those documents would have been refused, and…

  35. // note neural-bridge

    Reload the daemon when outbound_guard.py or fleet_heartbeat.py change

    Since #167 the Discord daemon imports scripts/outbound_guard.py, and every GitHub action passes through it. main.py has long imported scripts/fleet_heartbeat.py. The auto-reload watcher’s daemon-relevant list covered neither file. A change to either was pulled and then sat on disk until something else restarted…

  36. // note neural-bridge

    Block every direct write under knowledge/ except knowledge/agents/

    hooks/guard_concepts.py blocked direct tool writes only to knowledge/concepts/ and knowledge/quarantine/. An agent could still edit the tracked knowledge/connections/, index.md or log.md with its Write or Edit tool. That text then went public on the next manual commit without passing the filing gate or the…

  37. // hardening gemma-grc

    Gemma GRC: a fail-closed gate on what the corpus may contain

    The ingestion gate used to recognise one marking convention. A notes vault accumulates several, so the gate is now fail-closed. A note with any of these never enters the corpus:

    • a classification label other than public;
    • a private or confidential flag or tag;
    • a handling banner near the top;
    • a Classification: header line.

    Two details mattered more than the rules. A note saved with Windows line endings would have defeated the frontmatter match and walked straight past the labels, so line endings and malformed frontmatter are normalised before the gate reads anything. And banner matching has to tell a marking from a topic: “Restricted transfers” and “Confidential computing” are headings, not labels.

    Every chunk now carries a provenance class. Only public text and my own writing may reach a cloud teacher, and the cloud path is fail-closed: a corpus without the flag sends nothing.

    Chunk IDs are content hashes, so a re-ingest cannot silently re-point evaluation labels.

    The search index is stamped with the corpus and the model that built it. Loading an old index against the wrong embedding model used to fail silently. It now refuses, with a cosine of 0.59 as the evidence.

    There are 26 tests. Turning the classification check off makes six of them fail. An independent scan, written separately from the gate, finds no marked note in the build.

  38. // note neural-bridge

    Fix auto-reload's silent exit so merged work reaches the daemon; stop tests writing to the live vault

    The auto-reload watcher has exited 1 on every tick on main since 2026-05-13, logging nothing. #162 and #163 (merged 2026-09-23) never reached the running daemon; the Mac’s checkout is 17 commits behind origin/main.

  39. // note neural-bridge

    Skip flush for compile gate calls and for flush's own extraction call

    Closes the decision flagged on #162: compile-spawned sessions no longer trigger a flush model call. While fixing it, a second and worse instance of the same defect surfaced and is fixed by the same opt-out.

  40. // note neural-bridge

    Memory layer repair: silent failures fixed, wiki read loop closed, stores consolidated, pytest CI

    Steps 1 to 4 of the September build audit. Step 1 fixes the defects that made the memory layer fail silently and puts the suite under CI. Step 2 closes the wiki’s read loop, open since 2026-05-10. Step 3 executes the parts of docs/MEMORY_CONSOLIDATION.md that do not depend on Mac-side telemetry. Step 4 is the docs…

August 2026 · 14 entries

  1. // note neural-bridge

    Config that only existed in the scheduler

    Nothing in this repo read a .env file. Every environment variable the Telegram daemons needed was set inline inside its launchd plist, which had worked fine for a year because the only thing that ever ran them was launchd.

    Then I ran the check-in by hand. It exited with “no LUNA_TELEGRAM_ALLOWED_USERS configured” while the 07:40 job, running identical code down to the line, sent perfectly. Config that exists only inside one launcher is config you cannot test, and I had not noticed because I had never needed to.

    The loader is about eighty lines of standard library, matching how the rest of this codebase avoids dependencies for small jobs. The one design decision that mattered: anything already in the environment wins. A plist that still sets a variable keeps overriding the file, so adding a value could not change the behavior of anything already running. That property is what made this safe to do on a live machine.

    It loads from main() rather than at import, deliberately. The test suite imports these modules, and import-time loading would pull my real secrets file into the test environment and make results depend on which machine they ran on.

    Two things I did not know going in, both discovered the hard way.

    The luna-telegram entry in ~/Library/LaunchAgents is a symlink into the repo rather than a copy. So editing that plist edited a tracked file, and PlistBuddy helpfully reformatted the whole thing. Harmless once I diffed the parsed keys against the previous version, but not what I expected from a one-key deletion.

    And I took Luna’s Telegram bridge down for about three minutes. The unload succeeded, the reload failed with an I/O error because the unload had not finished yet, and I did not catch it until I checked the process list rather than the exit code. Retrying brought it back. The lesson is the ordinary one: bootout and bootstrap are not atomic, and a failed bootstrap looks a lot like success if you only read the first line.

    Also removed summarize-weekly from the installer. Its consumer was retired earlier the same day and the local job unloaded, but the installer would have cheerfully reinstalled it on the next run. Deleting a feature means deleting the thing that reinstalls it.

  2. // hardening neural-bridge

    Deleting a memory layer on the evidence

    The weekly lessons digest was supposed to compress each week into what the agents should carry forward. It ran every Monday at four, wrote a file per agent, and every agent prepended the newest one to every turn.

    I instrumented it two weeks ago. Over its measured lifetime it resolved for one request in seven, because exactly one agent ever had a digest directory. The other six turns in seven paid a four thousand character budget slot to look up a file that was not there, and got back an empty string that was indistinguishable from success.

    So it is gone. The one real digest went into that agent’s notes first, into the durable region rather than the session log, because it held her own correct diagnosis of the calendar bug I fixed yesterday. She had written it a week before I confirmed it. That was worth keeping; the machinery around it was not.

    The reclaimed budget went straight back out the door, which is the part I did not expect. Her notes file had grown to 17.6k against an 8k cap, and the section-aware budgeter was doing exactly what it was designed to do: drop the rolling log first, then start eating durable content. It was losing about 2.7k of the durable half on every single turn, including part of the section that holds the rules I have explicitly given her. Durable-first ordering was correct. There was simply not enough room for it to matter. The cap is now 12k, funded entirely by the digest removal, so the per-turn prompt is smaller than it was that morning.

    The general lesson is not that the digest was a bad idea. It is that I could only delete it because I had measured it. Before the telemetry existed I would have argued about whether it was useful, and the honest answer is that neither of us could have known.

  3. // fix neural-bridge

    The agent that said it could not, and could

    I asked for a summary of my day. She told me she could not reach the calendar or the inbox in that session. I ran both commands by hand about thirty seconds later and they worked fine.

    This is a fabrication, and it is the more expensive of the two directions. Inventing a meeting wastes a minute of my time. Inventing a limitation loses the task outright and quietly teaches me the assistant is less capable than it is, which is the kind of lesson that compounds. I had spent the previous day removing the opposite failure, a charter describing tools that did not exist. This was the same untruth pointed the other way.

    I tried to write my way out of it three times. A rule saying to attempt a tool before declaring it missing. A section naming the two surfaces she actually runs on, since she had claimed Telegram belonged to a different agent. Adding Bash to the tool list at the top of her Tools section, which had omitted it while the next paragraph told her to run CLIs with Bash. That contradiction was real and worth fixing regardless.

    It went from always to sometimes, and then it stopped improving. On the phrasing that triggered it hardest she still refused five times out of five.

    That is the signal that prompting was the wrong instrument. A model deciding whether it is permitted to look is a decision that should not exist in the first place.

    So the state now arrives already fetched. Today, the next seven days, and the threads I am waiting on get pulled before the turn starts and sit at the top of her prompt. She is not asked whether she can reach the calendar. The calendar is in front of her. The same prompt that failed five times now answers with real data three times out of three.

    Cached for five minutes, because conversation comes in bursts. Never blocks a turn, and a broken inbox still leaves her the calendar. When a fetch genuinely fails, the block says so in words she can repeat to me, which is the honest version of the sentence she had been inventing.

    Part of the cause was mine. A lesson I had migrated into her notes that same morning told her to distinguish “not authorized” from “not present in this surface.” Reasonable diagnosis, except I had given her no step that said to test first, so she was classifying failures she had never actually observed. It now reads: run it, then classify what the error says.

  4. // feature neural-bridge

    The tools she was documented to have, and never had

    Luna’s charter said she owned my calendar and inbox. Her allowlist hand-enumerated eighteen connector tool names for Calendar, Gmail and Drive. I tested it and she answered NO_CALENDAR_ACCESS.

    They had never worked. Auto-approving a tool does not create it, and app-level connectors are not loaded into a headless session at all. The charter had been describing a capability that did not exist, which is the same class of untruth as a monitoring dashboard reporting green over a dead process.

    The replacement is two read-only CLIs plus a scoped shell, which is the pattern already working for the career agent. Standard library only: the rest of the codebase reaches HTTP with urllib, and this needs a token refresh and two GETs. A forty megabyte SDK was not worth it.

    No send, no delete, no event creation. Not “not configured yet”, absent. Gmail is draft-only by charter, and Google has no draft-without-send scope, so the request stays read-only and there is no write path at all. The old allowlist had granted event deletion with nothing authorizing it. Tests assert this at the argument-parser level, so adding a send command breaks the suite rather than passing review.

    The most useful command is the one that lists threads I sent where nobody replied. On its first real run it surfaced a recruiter thread that had been silent nine days.

    Two setup traps, both of which I wrote into my own instructions before hitting them. Internal audience is offered whenever the project sits under an organization, but the organization attaches to the project rather than to a consumer identity, so consent fails at the last step. And Testing publishing status works the day you set it up, then expires refresh tokens for sensitive scopes after seven days, which would have taken her dark every week. Both are now documented with their exact symptoms.

    The first live conflict report flagged a flight against itself: the same flight sits on the calendar twice, once in English and once in Korean, occupying identical minutes. Duplicates are now told apart from clashes by requiring identical start and end plus a shared token that survives translation, such as a flight number. Two genuinely different meetings in one slot stay a conflict.

  5. // feature neural-bridge

    An assistant that speaks first, and watches the right thing

    Luna could only ever answer. Every surface she had required me to start the conversation, and the Discord fleet went ten weeks without a single mention because of exactly that. An assistant who never speaks first is a tool you have to remember to pick up.

    She now has two scheduled check-ins. The load-bearing design decision is that staying silent is a first-class outcome: a check-in that fires daily regardless becomes noise, noise gets muted, and a muted assistant is strictly worse than none. Silence is recorded as a success, because deciding there is nothing worth saying is different from failing to run, and the canary watches for the difference.

    The first version opened with fleet health. Agent uptime is devops wearing a persona, not executive assistance. What an assistant actually watches is the commitment board, so that is now the primary context and fleet health appears only when something is actionable.

    The board turned out to be carrying two regulator-escalated items seventeen days overdue while Luna had been silent for twelve weeks. That is the entire gap, in one line.

    Parsing it correctly took two attempts. The kanban date trigger is usually the date a card was ADDED, not when it is due, and the real deadline is written into the card text. Reading the trigger as a due date marked a card saying “Prep for the regulator’s first annual review by 2028-01” as thirty-three days overdue, and produced twenty false alarms on a board of twenty-two. An assistant that cries wolf on its first run is muted by its second. The trigger is not used consistently either, so the rule is now that a future date cannot be an added date, and an explicit in-text deadline always wins.

    Verified live: she leads with the escalated items, names the deliverables, recommends drafting both that morning, and says explicitly what can wait.

  6. // hardening neural-bridge

    Per-stage memory telemetry, and a canary that asserts on success

    Three memory layers were found broken, each having failed for weeks while every log line read healthy. The common shape was not a missed exception. A healthy layer and a dead one produced byte-identical output: nothing. Alert-on-error cannot catch that, because there is no error.

    So the instrumentation records every write, retrieve and utilize with store, agent, size and reason. Per stage, because the same visible failure needs a different repair depending on which stage broke. It records successes too, not only failures, since a store with zero events logged is indistinguishable from a healthy quiet one, which is exactly how a dead capture path survived ten weeks.

    The canary reads that log daily and asserts on the presence of recent success rather than the absence of errors. It names three shapes separately: SILENT (logged nothing while the fleet was active), FAILING (attempted, never succeeded), and DEGRADED (succeeds sometimes, below a rate floor). A dormant fleet reports IDLE rather than degraded, because this fleet genuinely sits quiet for weeks and a canary that fires on every quiet day gets muted, and a muted canary misses the real outage.

    Then the first fortnight of data showed Honcho failing 60% of writes, which turned out to be my own instrumentation firing during test runs and writing the suite’s mocked failures into the production log. Telemetry that lies about the system is worse than none, and it is the precise failure this whole effort exists to prevent. The recorder now no-ops under a test runner. The same pass found the canary too lenient: it only flagged total failure, so a store failing most of its writes still reported healthy because one landed.

    On clean data it correctly flags the weekly lessons digest at 1 in 7. That layer serves a single agent while costing a 4,000 character budget slot on every other agent’s turn.

  7. // release neural-bridge

    Four copies of the same pipeline, collapsed into one

    Every surface that talked to an agent ran its own copy of the same sequence: read the mention template, load the charter, build the prompt, get or create a session, look up tools and timeout and effort, call the model, retry once if the resume failed, touch the session, truncate the reply. Four copies existed, across the Discord daemon and three Telegram bridges, and the comments in them admitted it out loud: “mirrors handlers.py pattern”, “Mirrors luna_bridge.py for plumbing”.

    Copy number four is how drift starts. Adding a single per-agent effort flag the day before meant threading the same change through three files by hand, and one of them was missed on the first pass.

    run_agent_turn is now the only copy. Net 174 lines deleted. Two knobs keep the council room honest rather than forcing a fifth variant: a model override, and a stateless mode where each turn is rebuilt from the shared transcript so there is no session to resume and a failure must not trigger the retry.

    What stayed with each transport is what genuinely differs: chunking, structured action and attachment parsing, delivery, and the memory-capture call, which sits after a confirmed successful send so it never fires for a reply nobody saw. Discord takes the untruncated output because it parses action blocks before anything is trimmed.

    Twelve new tests cover the behaviors that used to live in four places: retry exactly once, never retry a fresh session, stateless skips the session store entirely, prefix prepending, and the model override. The live daemon was restarted onto the refactor rather than trusting the suite alone.

  8. // hardening neural-bridge

    Scoping an agent's shell before granting it

    Luna needed shell access to reach her calendar and inbox CLIs. Granting Bash grants the entire shell, which is a far larger permission than “read my calendar”, so the constraint shipped first and the grant second. The permission was narrow from its first minute rather than retrofitted later.

    The distinction that matters: --allowedTools auto-approves a tool, it does not constrain what the tool is asked to do. Bash on that list means any command. A PreToolUse hook returning exit 2 is the only mechanical enforcement point.

    It has to be per-agent, because hooks fire for every headless session launched from the repo. One allowlist shaped for Luna would have broken the career agent’s database CLI and the loop engineer’s test runner. The daemon now stamps an agent id into the subprocess environment, in exactly one place: a call site that forgot it would silently hand an agent the whole shell.

    Chaining, redirection and substitution are rejected before matching. Without that, appending a semicolon and a destructive command to an allowed one passes a prefix check. Fourteen of the thirty-two tests are attempts to defeat it that way.

    The fail-open is stated rather than hidden: an unstamped process is an interactive human session and is not constrained, because constraining it would break ordinary work in the repo.

    Verified through the real harness rather than only in tests. Asked to run ls /tmp as Luna, the command was blocked, never executed, and the model was told why. Asked to run her own calendar CLI, it reached the CLI, which then failed closed with a clear message and a pointer to the setup doc.

    The same commit removed eighteen hand-enumerated connector tool names from her allowlist. None of them had ever resolved. A test now fails if any agent lists one again.

  9. // release gemma-grc

    Gemma GRC: retrieval beat fine-tuning, and the search model was the lever

    The local compliance assistant shipped with no fine-tuned generator at all: base Gemma 3 4B, hybrid retrieval, and a prompt that demands a citation for every claim or a plain statement of what is missing.

    That was a measured decision, not a shortcut. Closed-book fine-tuning on my notes raised fabrication from 20% of answers to 48% on the 1B model. It learned the register and none of the facts. Retrieval raised answer quality by 0.86 on a five-point scale (confidence interval +0.48 to +1.24). Tuning the generator on top of retrieval added +0.04, which is noise.

    The component that moved was the one I had been ignoring. Contrastive fine-tuning of the 33M-parameter embedding model, 1,549 pairs, nine minutes on the Mac Mini, took recall@4 from 0.711 to 0.786. Four generator experiments bought nothing; one retriever run bought seven and a half points.

  10. // hardening neural-bridge

    Making silent failures loud: Honcho capture, in-flight mentions, dead links

    Three fixes in one day, all the same shape: something had been broken for weeks while the code looked healthy.

    The Neural Bridge to Honcho capture path had been dead from May 27 to Aug 2. Failures were swallowed at debug level, so one agent sat frozen at ten stored messages while every log line said fine. Failures now surface at warning, and the first successful submit in a process logs at info, so the logs positively confirm the path works instead of merely not complaining.

    A bot restart twelve seconds into a handoff killed the in-flight call with no reply, no retry, and no trace. Mentions are now registered in scripts/.inflight_mentions.json and cleared on completion; anything left over at startup gets a short notice posted to the affected channel by the owning agent, so the person who asked knows to re-send.

    Plus fourteen broken-link lint findings cleared and the Honcho integration docs refreshed to match reality.

    The pattern is worth naming, because it showed up three separate times in two days across three unrelated subsystems. Degrading to a quiet no-op keeps a daemon from crashing, and it is the correct local decision every time. Stacked together, they build a system that can lose most of its function and never say a word. Log the degradation.

  11. // feature neural-bridge

    recall.py: local semantic search over sessions, decisions, and daily logs

    Level 3 of the memory framework was deferred for months and is now filled. recall.py indexes vault Sessions, Memory, Daily notes and Meetings, the repo’s knowledge/, daily logs and decisions, and Claude Code’s own memory store, then answers semantic queries over all of it from one CLI.

    It runs entirely on the machine: ChromaDB with a bundled ONNX MiniLM, the same local-only stack already used in Bellwether. Nothing is sent anywhere, which is the whole point given the index spans meeting notes and personal daily logs. Indexing is incremental against an mtime manifest, and near-empty template sections are filtered and deduped at index time so a folder of half-filled note templates does not drown the results. The database under data/recall/ is gitignored and stays machine-local.

    Queried through a user-level skill, so asking what was decided about something reaches three weeks of context that no longer fits in a conversation.

  12. // note neural-bridge

    Council bridge committed, and a rule that contradicted itself

    The council bridge puts both advisors, one on Claude and one on a separate runtime, in a single Telegram room with a cheap router deciding who should answer each message. Twenty kilobytes of finished code had been sitting untracked for three weeks, one accidental delete from gone, so it is committed now. Committing is not activating: it still needs the manual bot setup steps, and the launchd job stays out of the installer until those are done.

    The other half is smaller and more embarrassing. There is a hard no-em-dash rule for everything these agents write, stated plainly inside the persona file. The governing documents around it contained nine em-dashes in one charter and ten in another. The rule was arriving alongside twenty counter-examples on every single read. The sweep script now takes a path so it can run against any agent’s files, with an exclusion for the one legitimate case: the line that states the rule has to be able to show the character.

    Output was already clean. The inputs were not, which is a reminder that context is not only what you tell a model, it is everything you hand it while telling.

  13. // fix neural-bridge

    Luna's memory was discarding her own rules to keep a changelog

    Luna keeps a working-memory file in the vault that the daemon injects into every prompt. There is an 8,000 character cap, and the injector kept the tail of the file on the assumption that a notes file is a log where the newest entries sit at the bottom.

    It was not a log. It had grown to 16,089 characters and was organized as a constitution with a changelog stapled underneath: standing preferences, voice notes, and a section titled “Decisions Andy has made that I should honor” at the top, an append-only session log at the bottom. Keeping the tail meant every turn handed her the changelog and threw away every rule above it. She had been reading a list of which pull requests merged in May instead of the sixteen decisions she was supposed to honor.

    The fix is a section-aware budget rather than a bigger cap, because a bigger cap fails later and just as quietly. Rolling-log sections are dropped first and durable ones are never dropped, and any trim at all now emits a warning. That warning is the actual repair. The arithmetic was a bug; the silence was the defect, and every layer of that memory stack was written to return an empty string on failure, which is how a system can lose most of its memory and never once say so.

  14. // feature neural-bridge

    Per-agent effort levels: stop paying for reasoning nobody asked for

    The Claude 5 generation models think by default, and effort is a dial with five positions from low to max. The daemon was not setting it, so every turn ran at the default depth. Asking an assistant what is on the calendar was allocated the same reasoning budget as a threat model.

    The fleet now carries an effort policy, one line per agent: low for conversational work and routing, medium for drafting and editing, high for research, security review, and PM triage. The autonomous coding loop runs at high and is the only thing in the house that has earned it. Invalid values are dropped rather than forwarded, because the CLI would warn and silently fall back, and a typo should not quietly change behavior.

    Verified against the live path rather than assumed. Worth noting for anyone routing a fleet through a proxy: claude-opus-5 is advertised by the Copilot endpoint but fails there with an assistant-prefill error, so the obvious upgrade would have taken every agent offline. The fleet stayed on 4.8.

July 2026 · 8 entries

  1. // note neural-bridge

    Agent vault homes consolidated, and a documented Bash exemption

    Agent home folders moved under a single Agents/ parent in the vault, with casing normalized so the directory names actually agree with each other. The charters, the daemon path constants, and the folders on disk are now consistent, which they were not: one agent’s charter pointed at a path that had been renamed underneath it.

    Separately, a policy test that asserted no agent may hold Bash in mention mode had been failing since the career-strategist agent legitimately gained it. That agent runs a journal CLI as its core function, so the grant is deliberate and scoped. The test now carries a single documented exemption with the reason attached, rather than sitting red and teaching everyone to ignore a red suite. An assertion that is expected to fail is not a test, it is a comment.

  2. // feature neural-bridge

    The loop engineer: an agent that claims its own issues and opens draft PRs

    The squad-discussion pipeline has been filing GitHub issues for months, and nothing consumed them. scripts/loop_engineer/ closes that loop: a separate launchd daemon polls for agent-ready issues, claims one via an atomic GitHub label swap, implements it in a per-issue git worktree with a fresh claude -p session, and opens a draft PR. Worktree isolation matters because the daemon shares one clone with an auto-reload watcher, and a loop switching branches underneath it would corrupt that view.

    Three gates decide whether a PR opens, and all of them are computed by the daemon rather than self-reported by the agent: existing tests may not be edited or deleted, the diff must stay under a line cap, and the change must introduce no new test failures against a baseline measured on the clean checkout before the agent touches anything. That last one is deliberate. The repo baseline is not green, so an absolute pass would have rejected every issue forever. Budget ceilings per run, per day, and per attempt are the kill switch.

    It ran end to end twice. The second time it implemented a --status flag for its own queue, added a test for it, and opened PR #161.

  3. // note neural-bridge

    Document post-PR `git checkout main` step for agent + me workflows (#112)

    Autonomous implementation of #112.

  4. // feature neural-bridge

    Multi-vote filing gate + calibration harness

    First improvement from the Fugu research: the filing gate now runs as a conservative multi-vote ensemble instead of a single judgment, plus a calibration harness to measure it.

  5. // note neural-bridge

    Loop_engineer: add a --status flag to report queue depth without claiming (#160)

    Autonomous implementation of #160.

  6. // note neural-bridge-blog

    Add CLAUDE.md with content contract and publish flow

    All five Zod collections (README documents only three), the Monday draft-flip publish flow, [skip-tweet], sync-paper slug caveat, and hard rules (cron owns draft:false, never rename published slugs). Also found sync-buildlog.yml is fully implemented, not a placeholder.

  7. // note neural-bridge

    Truth pass, reconcile README/AGENTS/STATUS with reality + roster drift check

    The self-documentation had drifted badly behind the code: README claimed 9 agents (14 defined), AGENTS.md claimed the V1 scaffold (3 agents, empty hooks), STATUS.md stopped at 2026-05-08.

  8. // note neural-bridge

    Mechanical write barriers, plugin skills, Sonnet 5 re-baseline

    • PreToolUse hook blocks direct writes to knowledge/concepts/ and quarantine/ (9 tests): the AGENTS.md rule is now enforced, not requested - First plugin-level skills: filing-gate-review, lint-triage; plugin 0.9.0 - Model pins re-baselined claude-sonnet-4-6 -> claude-sonnet-5 (flush, compile, lint, summarize_weekly) -…

May 2026 · 93 entries

  1. // feature neural-bridge

    Synapse DB ingester for vault corpus

    New scripts/echo/ingest_synapse_db.py walks the journal_entries table in ~/Development/Synapse/data/agent_i.db (Synapse’s career-intelligence SQLite) and produces a verbatim-cited markdown corpus at ~/Documents/Luna Master/Andy Profile/synapse-journal.md. Echo picks it up on her next mention via her existing…

  2. // note neural-bridge

    Honcho peer-memory + Luna Telegram bridge

    Two coupled changes shipped together: shared cross-agent peer memory via Honcho, and Luna reachable on Telegram as a second transport. They’re bundled because Luna’s Telegram bridge imports honcho_client — splitting them would mean a broken-import intermediate state on main.

  3. // feature neural-bridge

    Career strategist agent on Telegram with Synapse DB tool

    Loid is a new NB agent: a Hermes-pattern persona (charter + persona file + vault folder + handoff workflow) running on Neural Bridge. Reachable via a dedicated Telegram bot with Whisper voice ingest and via @loid in Discord. He’s the conversational layer over Synapse’s career-intelligence database, sibling to Yor…

  4. // fix neural-bridge

    Charter now matches allowlist; can ship daemon config from Discord

    Luna’s allowlist (scripts/discord_bot/repos.py:58) has had {neural-bridge-blog, neural-bridge} for a while now. But her charter (plugins/neural-bridge-core/agents/luna.md line 107) still says:

  5. // feature neural-bridge

    Expand per-agent response caps (originally proposed by @luna)

    Originally proposed by @luna in Discord on 2026-05-11 evening; she couldn’t ship it herself because of the charter-vs-allowlist drift fixed in #132. Shipping by hand to clear her stranded working-tree change.

  6. // note neural-bridge

    Draft "The 6 layers" blog post into Obsidian-mirrored drafts

    First public blog post on the Neural Bridge spine (V2 step 4 from docs/STATUS.md). Per-comment direction: full-prose draft only, hold on publishing live, file into the Obsidian vault, downplay the Karpathy / Cole Medin lineage.

  7. // fix neural-bridge

    Align 5 agent charters with push allowlist (follow-up to #132)

    Follow-up to #132. Audited the other five agents in the push allowlist for the same class of charter-vs-allowlist drift that bit @luna. Found three real gaps and two minor cleanups.

  8. // fix neural-bridge

    Correct model pin from claude-sonnet-4-7 to claude-sonnet-4-6

    Closes #135. One-line fix to align senior-pm with the rest of the agent roster (all use claude-sonnet-4-6).

  9. // note neural-bridge

    Consolidate never-shell-fallback rule into the shared prompt

    Closes #137. Final follow-up from #134 and senior-pm’s triage report.

  10. // feature neural-bridge

    Weekly profile synthesis cron — close the loop between raw corpus and structured files

    Andy’s question this evening: “is Echo actually building memory and corpus from all of our interactions and conversations in the neural-bridge discord server?”

  11. // feature neural-bridge

    Wire claude-transcripts ingestion into a daily cron (closes #143)

    Closes #143. Companion to #142.

  12. // feature neural-bridge

    Ingest MindFrame Discord conversation logs (closes #144)

    Closes #144. Third Echo input pipeline alongside raw-conversations (Neural Bridge Discord, #142) and claude-transcripts (Claude Code sessions, #143).

  13. // hardening neural-bridge

    Auto-checkout main after open_pr_with_changes push (closes #126)

    Replaces #126 (Copilot’s docs-only PR), which was blocked: its proposed wording claimed the daemon auto-runs git checkout main after open_pr_with_changes, but the daemon did not yet do that. This PR ships the implementation and the docs together so the assertion is true on merge.

  14. // fix neural-bridge

    Spell Andrej Karpathy correctly

    One-character fix surfaced during editorial review of PR #128: ATTRIBUTION.md line 7 had “Andre Karpathy” where it should be “Andrej Karpathy” (the same typo PR #128’s draft inherited and already fixed on its own branch).

  15. // feature neural-bridge-blog

    Rewrite Korean translations in Yozm-Wishket blog voice

    • Rewrites all 20 Korean sidecar files in publication-grade Korean blog voice - Grounds the translation prompt in real Korean tech-publication samples (요즘IT / Yozm Wishket primary, IT World Korea + JoongAng IT for terminology) - Output should now be indistinguishable from hand-written Korean blog prose
  16. // feature neural-bridge-blog

    Per-article translation toggle, single English URL per page

    Replaces the sidecar /ko/* route mirror pattern with an inline translation toggle. One canonical English URL per article; the Korean translation renders into the same page (hidden by default) and is revealed via a per-article toggle button.

  17. // fix neural-bridge-blog

    Replace INFO 310 with Cybersecurity & Enterprise Risk Management

    Updates the Adjunct Lecturer line in the about-page profile block.

  18. // feature neural-bridge-blog

    Korean translation + bilingual toggle on about page

    Adds Korean to the about-me page using the same inline-toggle pattern as PR #26 (research / posts / projects / agents). About-me is a hardcoded Astro file, not a content collection, so the bulk pipeline doesn’t cover it. Hand-translated following the Yozm-Wishket voice prompt.

  19. // fix neural-bridge-blog

    Skip Korean sidecars + dedupe against existing open issues

    Today’s content-cleanup wind-down generated 12 spurious tweet-draft issues because the tweet-on-publish workflow re-emitted drafts every time an already-published article was modified in place (em-dash sweep, editorial pass) and treated Korean sidecar files as separate publishes despite the inline-toggle…

  20. // note neural-bridge-blog

    Add prompt-injection-as-compliance-risk draft for May 18 publish

    New research draft commissioned from the content agent. Picks up item #3 from the GRC x AI Security backlog: “Prompt injection is a compliance risk, not just a security bug.”

  21. // note neural-bridge-blog

    V2 ships completion milestone post for May 25

    New post for the May 25 publish slot. Closes the V2 arc with a milestone post that does three things:

  22. // note neural-bridge-blog

    Ship V2 milestone post today (pubDate 2026-05-10, draft: false)

    Bumps the V2 milestone post forward from the May 25 cron slot to ship today. Pairs with the Korean sidecar generated by the translation pipeline this session so the 한국어 toggle works on day one.

  23. // note neural-bridge-blog

    Add Korean sidecar for prompt-injection-as-compliance-risk

    Generates the Korean translation for the June 1 prompt-injection article ahead of time so the 한국어 toggle works on day one and we avoid betting on the May 31 Sunday-prep cron hitting a clean claude -p invocation.

  24. // note neural-bridge-blog

    Rewrite V2 ships Personality paragraph

    Andy’s edit on the live V2 ships post. Drops the 애교 / 존댓말 mechanics from the Personality bullet and reframes the change as ‘Luna sounds more human and less structured’ without exposing the Korean-specific internals.

  25. // feature neural-bridge

    Tighten dirty-tree check to file-intersection only

    Today’s Luna debug session hit the daemon’s dirty-tree check twice: - First on a stale modified src/pages/about.astro (cleared by neural-bridge-blog#27) - Then on an untracked public/images/andy-profile.jpg (cleared by neural-bridge-blog#28)

  26. // fix neural-bridge

    Reap claude -p subprocess on timeout (closes #110)

    Closes #110.

  27. // note neural-bridge

    Escalate silent staleness after N skips (closes #111)

    Closes #111.

  28. // feature neural-bridge

    Auto-add agent-inbox issues to V1→V2 project board (closes #112-adjacent gap)

    Closes the systemic gap noticed in this session: issues filed with the agent-inbox label weren’t landing on the Neural Bridge V1 to V2 project board. #110, #111, #112 all had projectItems: [] until I added them manually.

  29. // feature neural-bridge

    Accept .pptx and .xlsx — Luna can read Office docs end-to-end

    Luna told Andy she couldn’t process Office docs from Discord. .docx was already on the allowlist (#103, #104), but .pptx and .xlsx were rejected as unsupported_extension — that’s what she actually hit. Probably overgeneralized into “I can’t process office docs” in chat.

  30. // feature neural-bridge

    Claude session resumption per (channel × agent) — Luna remembers across turns

    Solves the user-reported “Luna can’t remember what we said earlier in this thread” problem. Screenshot evidence: mid-NIS2-thread, Luna admitted file content from prior turns hadn’t carried forward and her Discord history visibility cut off mid-sentence.

  31. // feature neural-bridge

    Per-agent Discord conversation archive in Obsidian vault

    Third memory layer for agents, alongside session resumption (#119) and Luna’s notes.md. Every Discord turn is appended to a markdown file in the agent’s vault subpage. After the session TTL expires and claude’s in-memory state is gone, the markdown is what’s left for the agent to consult.

  32. // feature neural-bridge

    Expand push allowlist + 50-msg history + Luna 11400 cap

    Three related tweaks to how agents handle Discord chat.

  33. // feature neural-bridge

    Weekly auto-summarization → per-agent lessons-learned digests (closes #121)

    The fourth memory layer. Compresses the verbatim conversation archive (#120) into a per-week “lessons learned” digest per agent. Auto-injects into every mention prompt the following week. This is the part that distills raw turns into “Andy prefers X” abstractions without manual curation.

  34. // feature neural-bridge

    Cross-agent shared conversation log in guild channels (closes #122)

    Closes the cross-agent memory gap from the four-layer roadmap. After #120 each agent had a private conversation archive — fine for DMs, wrong for guild channels where multiple agents naturally participate in the same thread. Luna couldn’t see what Echo said in #neural-bridge even though Andy could; Echo couldn’t see…

  35. // feature neural-bridge

    Semantic search via Ollama bge-m3 + sqlite-vec (closes #123)

    Adds the embedding index over the per-agent conversation archive (#120) and a search_conversation_memory action. Agents can do semantic recall, not just literal Grep.

  36. // hardening neural-bridge

    True 24/7 uptime: caffeinate wrapper + auto-reload watcher + rotating logs

    Three reliability moves on the Discord daemon. caffeinate -s -i wraps the launchd entry so the Mac never silently sleeps the daemon’s process tree even with pmset already configured. An auto-reload watcher runs every two minutes: git pull main, and if anything new landed, the daemon restarts itself. A rotating-file logger caps individual log files at 10 MB with seven backups, ceiling around 70 MB total, no more unbounded log growth on the Mac Mini.

  37. // feature neural-bridge

    Discord DMs treated as implicit mentions

    When you DM any of the agent bots directly, the message is treated as an implicit mention of that agent. Removes the friction of @-naming an agent in your own DM channel where there’s only one bot in the conversation. 1:1 chat with any specialist works the way you’d expect.

  38. // milestone neural-bridge

    First real wiki harvest: 7 concept articles, 21 connections, 8 quarantined

    The compile pipeline produced its first real output. Seven concept articles landed in knowledge/concepts/ with rich bodies, twenty-one connection files surfaced shared-session relationships, eight candidates hit the quarantine path with reasons. The substrate is no longer a scaffold; it’s a wiki that’s growing. Every concept carries provenance frontmatter, which session it came from, transcript sha256, source log, so anything later can be traced back to the moment it was decided.

  39. // milestone neural-bridge

    Librarian agent: vault index + audit + restructure proposals

    The 11th specialist. The librarian keeps the Obsidian vault organized: maintains an index, runs audits for orphan notes and dead links, and proposes restructures when sections get crowded. Closes a friction I’d flagged in the Phase B core notes, the vault was filling with SOPs, drafts, and journal entries with no agent to keep the shape coherent. Now there is.

  40. // feature neural-bridge

    Nightly compile job at 03:00 daily

    The compile pass now runs unattended at 03:00 PT daily under launchd. Reads the new daily-logs since the last successful run, runs each candidate through the filing gate, calls the rich writer for PROMOTEs, archives any prior versions, surfaces shared-session connections, refreshes log.md and index.md, posts the Discord summary. The wiki is now a self-growing system; I just have to do work, and the substrate compiles overnight.

  41. // fix neural-bridge-blog

    Responsive page widths — wide prop for index/grid pages

    Index/grid pages now opt into a wider container (max-w-6xl ≈ 1152px) instead of being forced into the article-reading width (max-w-2xl ≈ 672px). Article reading pages stay narrow for readability. Nav and Footer accept the same prop and stay aligned with whichever width the page uses.

  42. // feature neural-bridge-blog

    Collapse Posts / Research / Projects into a single Writing tab

    Top nav had seven items and was getting cluttered. Folds Posts / Research / Projects under a single Writing parent. New /writing hub page shows three sections (Posts / Research / Projects) with 3 most recent items each + an “All X (N) →” link with the total count. Direct /posts, /research, /projects routes…

  43. // feature neural-bridge-blog

    Korean translation infrastructure + Memory Poisoning paper translated end-to-end

    PR A in the Korean-translation feature. Lays the full infrastructure and proves it end-to-end on one flagship article so you can see the UX working before the bulk translation runs.

  44. // feature neural-bridge-blog

    Bulk-translate all articles + agents to Korean (PR B)

    Closes the bulk-translation phase. The site is now fully bilingual across all curated content.

  45. // note neural-bridge-blog

    Queue Memory Poisoning sequel + Six Layers; promote Reg v0.2 in-place

    Three content moves that fill the publishing queue for the next two Mondays and update the live regulation post.

  46. // note neural-bridge-blog

    Rewrite to reflect live-deployed state

    Previous README claimed “V1 scaffold. Not yet deployed.” That hasn’t been true since the site went live. Mirrors the same kind of cleanup that landed in andy-herman/neural-bridge#63.

  47. // feature neural-bridge-blog

    Real chronological feed + 18 seed entries covering V1+V2 ship arc

    The /buildlog page has been a “coming in V2” placeholder since the site launched. Replacing it with a working chronological feed seeded with 18 hand-curated entries that cover the V1+V2 shipping arc through 2026-05-10.

  48. // feature neural-bridge-blog

    Auto-sync workflow + script — daily fetch of merged PRs (re-target to main)

    #7 was opened with feat/buildlog-feed as its base for the stacked-PR pattern. When #6 (the buildlog feed) was squash-merged to main earlier today, the base branch got deleted and #7 no longer has a…

  49. // feature neural-bridge-blog

    Roster page foundation — 13 agent cards + avatar pipeline + nav

    The /agents foundation PR. Index page with 13 cards, full avatar fetch pipeline, content collection, nav update. Detail page per agent ships in a follow-on PR.

  50. // fix neural-bridge

    Wire echo and ux-designer into MENTION_ALLOWED_TOOLS + ADD_DIRS_PER_AGENT

    Both plugin files existed but neither was in the per-agent tool allowlist or the per-agent add-dirs map — same bug class that broke Luna and librarian earlier (tools=none, add_dirs=0 on every mention).

  51. // fix neural-bridge

    Allow $TMPDIR as a second root so agents can attach Drive-downloaded files

    The validator in #96 only allowed paths under \$HOME. That blocks Luna’s natural flow: when she downloads a Drive file via download_file_content, Python’s tempfile.NamedTemporaryFile lands it at /var/folders/... on macOS (resolves to /private/var/folders/...) — outside \$HOME, so the attachment never went…

  52. // fix neural-bridge

    Remove local 'import discord' from handle_mention (UnboundLocalError on every mention)

    Andy @-mentioned Content Manager in #content-backlog at 12:57 PM. Bot stayed silent. Logs show:

  53. // feature neural-bridge

    Inbound attachment ingest — drop .eml/.pdf/.docx/.txt into chat

    Wires Echo to accept files Andy drops directly via Discord. End goal from the conversation: feed Echo emails, exports, PDFs, and plain text so she can build a sharper personality model. When she gets good enough, chatting with her should feel like talking to Andy.

  54. // feature neural-bridge

    Per-agent allowlist + drop dir — Luna accepts images

    • PR #103 wired inbound attachment ingest but hard-gated it to Echo with docs-only extensions. Dropping a PNG into Luna’s DM did nothing — daemon never saw the attachment. - This PR replaces the single-agent gate and single-allowlist with per-agent maps. Echo unchanged. Luna joins the table with docs + image formats…
  55. // note neural-bridge

    Operating Rule 10 — Korean translator skill

    Closes the loop on the bilingual blog. Content agent now owns Korean translation alongside its drafter role.

  56. // feature neural-bridge

    Sunday-prep auto-translates the upcoming post to Korean

    Closes the bilingual publishing loop. Every Sunday-prep run now generates a Korean sidecar alongside the LinkedIn variant and X draft.

  57. // note neural-bridge

    Agents push code to GH via open_pr_with_changes (chat-gated)

    Closes the gap where agents could draft blog posts and recommend changes but had no way to ship them. Luna’s complaint that triggered this: asked to push a fix to the blog, she correctly identified the gap but fabricated an “open a PR directly against GitHub” workflow that didn’t exist.

  58. // fix neural-bridge

    Scope blog --add-dir to src + public, skip node_modules

    PR #107 wired the full blog repo root into --add-dir for luna, content, and ux-designer. That includes node_modules (~14k files). Claude’s session startup does a directory inventory, which made Luna’s first post-#107 mention hang for 3+ min at ~2% CPU steady — matches “scanning a huge tree” not “doing inference.”

  59. // feature neural-bridge

    Add Personality section — light playfulness + 애교 within 존댓말

    Charter-only change. Adds a new “Personality and playfulness” section between Tone and Language.

  60. // note neural-bridge

    Rewrite to reflect V1-shipped state

    The previous README still described Neural Bridge as a “V1 scaffold — not yet functional” with three agents and an unimplemented hook pipeline. That hasn’t been true for ~50 PRs.

  61. // fix neural-bridge

    Define user_id in handle_mention scope (NameError hotfix)

    handle_mention references user_id in the MENTION log line:

  62. // note neural-bridge

    Idea-generation, editor-handoff, publish-readiness rules

    Three new operating rules added to the content agent’s plugin definition. Each closes a gap that surfaced while testing the new publishing pipeline:

  63. // feature neural-bridge

    Auto-update agents.json when create_agent carries client_id

    Closes a manual step that was slowing tonight’s Luna recruitment. When the recruiter’s create_agent action carries a client_id field, agent_builder.py now also appends the entry to scripts/discord_bot/agents.json so the daemon picks up the new bot on next reload — no hand-edit required.

  64. // fix neural-bridge

    Backfill luna entry into agents.json (lost in #75 squash-merge)

    PR #75 was meant to ship Luna’s charter + KNOWN_AGENTS + agents.json entry. The agents.json commit got dropped in the squash-merge (see #75 merged-files list — only 3 files). This backfills it. One-line content fix to wake Luna up.

  65. // fix neural-bridge

    Per-agent claude -p timeout (teaching-prep 600s, recruiter 480s)

    The 300-second default in claude_invoke is enough for typical mention flows but routinely insufficient for two specific agents:

  66. // fix neural-bridge

    NB_AGENT=compile + NB_NO_DISCORD=1 on spawned subprocesses

    Running compile.py against 70 corpus concept candidates exposed two related problems flagged by Andy in real-time:

  67. // fix neural-bridge

    Stop truncating content/social replies + mark every chunk + bump timeouts

    Content agent posted a draft summary in Discord. Message ended mid-sentence at “DORA practitioners will r…” with no continuation. Looked like the response was just cut off.

  68. // fix neural-bridge

    Wire luna and librarian into MENTION_ALLOWED_TOOLS

    Both agents were missing from the per-agent tool allowlist. Every mention against them passed tools=none to claude -p. Luna’s Calendar create_event call tonight was rejected silently because the MCP tool wasn’t in her allowlist.

  69. // feature neural-bridge

    Persistent memory via vault notes auto-injection + no-fabricate charter

    Two related fixes to Luna’s behavior, addressed in one PR because they share the same charter file.

  70. // fix neural-bridge

    Handle bootout race in install.sh

    1. “Bootstrap failed: 5: Input/output error” on the discord-bot agent. bootout returns to the shell immediately but launchd is still reaping the service in the background. bootstrap too fast (the original code did them back-to-back) hits a transient I/O state and fails. Hit Andy at least twice this quarter.
  71. // feature neural-bridge

    Add front-end-designer — visual designer for the blog and future web surfaces

    The 12th specialist agent. Owns the look and feel of neural-bridge.dev and any other web surface that gets built later.

  72. // note neural-bridge

    Add Rule 9 — buildlog-entry mode

    Closes the gap between the buildlog auto-sync workflow (fact-based skeletons) and the build-in-public voice that hand-curated entries carry. Rule 9 gives the content agent a clear playbook for both drafting buildlog entries from scratch and improving auto-synced entries that came out too thin.

  73. // feature neural-bridge

    Add echo — Andy's self-knowledge cataloger (Phase 1)

    Phase 1 of 5 in the Echo build. Charter + scaffolding for a new specialist that maintains a structured profile of how Andy thinks and writes. Other agents (Luna, content, social, professor) consume the profile to mirror Andy accurately.

  74. // note neural-bridge

    Echo Phase 3 (Discord accumulator) + ux-designer rename + agents.json wiring

    Combines Phase 3 of the Echo build with two tangential changes that got bundled into the commit history accidentally during a force-push. All three are correct; just easier to ship together than untangle.

  75. // feature neural-bridge

    Direct file attachments via `attachments` block (≤24MB)

    Closes the gap where Luna (and other agents) had to fall back to “open from this path” instructions because the daemon couldn’t post Discord file attachments. Agents can now emit a fenced ```attachments block at the end of their reply with a JSON array of absolute paths; the daemon validates + attaches via…

  76. // feature neural-bridge

    Post_as CLI — message as real bot identity

    Replaces the webhook-with-username-override hack from the Echo build phase pings. The webhook worked but posted from a webhook entity (wrong avatar, no bot-identity continuity). This posts AS the bot via Discord’s REST API + keychain token.

  77. // feature neural-bridge

    Drive-overflow charter rules + Drive Map convention

    Pairs with #96 (direct attachments ≤24MB) to close the file-sharing loop. Two new sections in luna.md, inserted between the vault layout and the vault-content-discipline section so they flow with the existing “where things live” framing.

  78. // feature neural-bridge

    Voice authentication review mode

    Echo now has two jobs: profile maintenance (existing) + voice authentication review (new).

  79. // feature neural-bridge

    Phase 4 — Claude transcript ingestion (opt-in whitelist)

    Phase 4 of the Echo build. Script that walks Andy’s Claude Code transcripts (~/.claude/projects/<id>/<session>.jsonl), filters out system-injected wrappers, and appends user turns to Andy Profile/claude-transcripts.md for Echo’s corpus.

  80. // feature neural-bridge

    Phase 5 — voice profile auto-inject into content/social/luna

    Final phase of the Echo build. Other agents now consume Echo’s voice profile when generating user-facing prose.

  81. // post neural-bridge-blog

    Research: AI Security Regulation in 2026

    Working practitioner’s map of the AI security regulatory landscape: EU AI Act, NIST AI RMF, US state-level patchwork, ISO/IEC 42001, plus how the sector overlays (HIPAA, DORA, NIS2) intersect. Not legal advice. The thing I needed when I was scoping what compliance actually changes about how a team builds.

  82. // feature neural-bridge

    compile.py Phase B core: rich concept writer + never-overwrite history

    A PROMOTE verdict now produces a real concept article instead of a slug-and-summary stub. Separate claude -p call after the gate decides, structured prose: intro, key points, how-we-use-it, open questions, related-concept wiki-links. Re-promoting an existing concept archives the prior version to concepts/.history/<slug>/<timestamp>.md first, audit trail the lint pass needs. Live runs append to knowledge/log.md and add wiki-links to knowledge/index.md. 47 tests passing.

  83. // feature neural-bridge

    compile.py Phase B expansion: connections, --agent, --flush

    Three additions. Connection writer: PROMOTE’d candidates that share a source session_id get a knowledge/connections/<a>--<b>.md file by construction. --agent flag: scope a run to one agent’s daily-logs (compartmentalization recommendation from the memory-poisoning paper). --flush: short-circuits to a single-session manual flush via hooks/flush.py, folded from the original flush.py issue. Closes V2 issue #9 end-to-end. 62 tests passing.

  84. // feature neural-bridge

    Dashboard script: dated markdown reports from gh data

    scripts/dashboard.py pulls GitHub state (open issues, recent PRs, label counts, kanban column placement) and writes a dated markdown report. Useful as input to senior-pm’s weekly summary and as a snapshot I can drop into the vault. No new infra; just structured output from the data that’s already there.

  85. // milestone neural-bridge

    Discord orchestrator goes live with nine specialists

    Nine bot identities, one daemon, one asyncio loop. Mention any agent in #neural-bridge and the right specialist responds. Agents can emit a structured actions JSON block to file issues, comment, label, close, no Bash, no shell, audit trail per call. Cross-agent handoff via @-mention propagation, capped at five turns per Andy-initiated thread to prevent runaway chains.

  86. // milestone neural-bridge

    Luna agent: executive assistant with calendar + Gmail MCP

    The 10th specialist. Luna handles calendar and email triage, different scope than the build-side agents. Wired into Discord with MCP servers for Google Calendar and Gmail so she can read upcoming meetings, surface conflicts, draft replies, and propose schedule moves. The first agent whose work isn’t strictly about Neural Bridge itself; the substrate is starting to get used.

  87. // post neural-bridge-blog

    Research: OWASP for AI

    The technical-control vocabulary that pairs with the regulatory landscape piece. OWASP didn’t pretend AI was a web app. They built a separate Top 10 for LLM applications and they’re already on the second iteration. Field guide for using OWASP as vocabulary, as self-assessment, and as regulatory backstop.

  88. // feature neural-bridge

    Professor agent: corpus-aware deep research, long-form output

    The teaching-prep agent gets rewritten as the professor: aware of the INFO 310A lecture and lab corpus, set up for long-form research and lecture-prep output. Came alongside scripts that extract structured content from the live deck and lab files so the agent reads what’s actually being taught, not what was hand-summarized last week.

  89. // feature neural-bridge

    Recruiter `create_agent` action: agents can spawn agents

    The recruiter agent can now provision new specialist agents end-to-end. Emit a structured create_agent action and the daemon writes the plugin file, updates KNOWN_AGENTS in two places, bumps the plugin version, branches, commits, pushes, and opens a PR. Manual Discord-side steps (token, application, invite) stay with me on purpose, those involve secrets that should not be automated.

  90. // release cross-cutting

    Publishing pipeline online: Sunday prep, Monday publish

    Weekly cadence wired up. Sunday 18:00 PT a launchd job picks the next queued draft, generates a LinkedIn variant via claude -p against my actual voice corpus, and posts a Discord brief listing the three artifacts (blog body + LinkedIn variant + X draft) for my Sunday/Monday review. Monday 18:00 PT a GitHub Actions cron flips draft: true → published; Vercel auto-deploys. Most of the plumbing was already wired before I noticed; the gap was the cadence and the LinkedIn voice generation.

  91. // post neural-bridge-blog

    Research: Memory Poisoning in Personal Agentic AI Substrates

    The threat model that drives every design decision in compile.py. AgentPoison and PoisonedRAG show that less than 0.1% adversarial entries in long-term memory can hit 80-99% attack success. If agents write into shared memory and other agents read from it, you have a perfect channel for prompt injection, and prompt-rule defenses do not survive contact with content the prompt itself loads.

  92. // milestone neural-bridge

    V1 scaffold lands: plugin marketplace, three specialist agents, wiki skeleton

    The first version of Neural Bridge ships as a Claude Code plugin marketplace. Three specialist agents (research, teaching-prep, content) live in plugins/neural-bridge-core/agents/, the wiki skeleton is in knowledge/, and ADRs cover the early design decisions. The hooks and compile pipeline are stubs at this point, the substrate is real but not yet alive.

  93. // post neural-bridge-blog

    First public post: Why I'm building Neural Bridge

    The blog goes live with the framing post. Neural was the first attempt at a substrate. Synapse refined the architecture. Argus turned the lessons into a regulatory-compliance copilot at work. Each one taught me something I needed for the next. The substrate didn’t compound. Until now.

// adding to the feed

Manual entries land in src/content/buildlog/<YYYY-MM-DD>-<slug>.md. Auto-sync from merged PRs across neural-bridge and neural-bridge-blog ships in a follow-on PR.