fix: re-enable history shortcuts by jakst · Pull Request #503 · anomalyco/opencode

jakst · 2025-06-28T14:36:38Z

Arrow keys up/down were not working for navigating through history. They were commented out in 568c047, looks like an accident, but I could be wrong.

Arrow keys up/down were not working for navigating through history. Looks like they were commented out in 568c047, looks like an accident, but I could be wrong.

adamdotdevin · 2025-07-01T10:33:46Z

not an accident, there were other issues, but those might be resolved now. however, currently working on a feature that likely supplants this, stay tuned

…on for any source (anomalyco#503) * feat(skill): add data-scraper-agent skill Workflow skill for building AI-powered public data collection agents. Covers any scraping target: job boards, prices, news, GitHub, sports, events. - Full architecture guide (config.yaml, scraper/, ai/, storage/) - Gemini Flash free tier client with 4-model fallback chain - Batch API pattern (5 items/call) — stays within free tier - Feedback learning loop from user decisions - Notion / Sheets / Supabase storage templates - GitHub Actions cron schedule (100% free) - Anti-patterns table, free tier limits reference, quality checklist - Real-world examples and reference implementation (job-hunt-agent) * fix(skill): address PR anomalyco#503 review violations in data-scraper-agent - Read batch_size from config.yaml instead of hardcoded constant - Branch main.py on storage.provider; label example as Notion-only - Replace undefined sync_feedback() with load_feedback() + comment - Add commented Playwright browser install step to CI workflow - Add permissions: contents: write; remove silent `git push || true` - Remove external unvetted repo link from Reference Implementation - Move import json to top of pipeline.py block (was after usage) - Guard context.md read with exists() check; fall back to empty string - Replace deprecated datetime.utcnow() with datetime.now(timezone.utc) - Remove duplicate config.yaml entry from project directory template

Ships the recent multi-session work now on main: #150147 push-retry reliability fix, iris agents update (--enable-integration/--add-tools), bloqs make-public/share, atlas:item publish-markdown + harden + image-upload, iris integrations pathways onboard (anomalyco#503), iris mcp register (#150264), outreach-strategy alias, learn/okf. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Wire the two primitives IRIS already shipped but never connected: on-device whisper.cpp STT (lib/transcription.transcribeLocal, already used by `ideas capture`) + the V6 ReactLoop chat brain (streamAgentChat). Adds the missing local TTS + push-to-talk turn-taking so you can hold a spoken conversation with any agent (e.g. TOBI anomalyco#642) fully on-device — $0/turn, offline-capable, HIPAA-safe. Cloud voices (ElevenLabs/VAPI) stay in `iris voice` for phone. - lib/voice.ts: captureMic (ffmpeg push-to-talk → 16kHz WAV), speak (macOS `say` default, Piper optional, never throws), listMics (device discovery). - chat: --voice loop with multi-turn conversation_history (no server session, no cross-turn poisoning); flags --mic / --tts / --tts-voice / --list-mics. Plan + tracking: bloq anomalyco#503 item #158044, bug #158045. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…tdin hang The push-to-talk loop hung on the SECOND stdin read ("press ENTER to stop"): manual `process.stdin.once('data')` + pause/resume fought the prompt lib's raw-mode TTY in Bun, so the stop keypress was never seen. Fix: - captureMic now auto-stops on silence (ffmpeg `silencedetect`, ending the turn on trailing silence after speech / a paused opening utterance), with a manual ENTER override via `stopSignal` for rooms where the VAD threshold misfires. - The loop uses ONE readline interface for every turn (question + rl.once('line') stop, removed after) instead of raw stdin juggling — works across turns. - Dropped the clack spinner inside the loop (it flips stdin to raw mode); status is plain stderr now. Blank/`[BLANK_AUDIO]` transcripts are skipped. Result: talk → pause (or ENTER) → agent replies + speaks → repeat. No hang. Tracking: bloq anomalyco#503 #158044, bug #158045. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…+ clean UI Field feedback: "works but not seamless." Three fixes: - Hands-free after turn 1: the agent answers, then the loop auto-listens again (no ENTER per turn). ENTER at start; two silent captures fall back to ENTER so we never hot-loop on room noise; say "goodbye"/"quit" (or Ctrl-C) to hang up. - Speakable replies: a per-turn voice hint steers 1-2 short spoken sentences, no markdown/lists (appended to the query only, not stored/displayed, so it's backend-agnostic — no dependency on a system role). Long paragraphs were the worst part of TTS. - Clean line-based output: dropped the \r status writes that collided with the readline echo (the stray ">"/"TER)" artifacts). Tighter 1.1s silence cutoff. Tracking: bloq anomalyco#503 #158044, bug #158045. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…can't hang Silence auto-stop was unreliable across rooms/mics (field report: "really bad at detecting silence", and it hung when the user didn't speak — silencedetect never fired and the ENTER override via rl.once('line') could miss the keystroke). Rewrite to a fully deterministic loop: - START and STOP are BOTH rl.question() reads (the start read always worked in the field, so reuse the exact mechanism for stop). ENTER records, ENTER sends. No keystroke can be missed; ffmpeg is SIGINT'd the instant ENTER resolves. - Silence detection is now opt-in behind captureMic({autoStop}); default off. - 60s hard cap as a safety backstop only. Short-reply voice hint retained. Result: predictable, hang-proof turn-taking that works in any audio environment. Tracking: bloq anomalyco#503 #158044, bug #158045. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…I (#165979) Three subcommands over the fl-iris-api /api/threads API (IRIS_API base): - message <agent> <content> --thread|--to : post into a room AS an internal agent (uses the as_agent_id keystone); --no-trigger suppresses replies; --to opens a fresh DM thread with the recipient auto-responding. - thread [id] : list threads, or show one room's participants + messages. - inbox <agent> : list rooms an agent participates in. Ensures the sender is an internal participant before posting (server rejects otherwise). buildThreadMessageBody() is pure + unit-tested. Surface half of the multi-agent rooms epic (bloq anomalyco#503 / list anomalyco#1688); pairs with fl-iris-api c4be6710. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Every data path we shipped pointed inward. `data-sources sync` pulls cloud storage INTO a bloq, `bloqs ingest` uploads a file INTO a bloq. The only way out was a per-entity `pull <id>` (boards/leads/agents/workflows/pages/products/ events), one id at a time, CLI-only — and `iris bloqs` had no pull at all, so the container itself (lists, items, attachments) could not be exported. `iris export` exists but exports a chat session, not workspace data. So "can I get my data out?" had no good answer, which is both a real data-loss exposure and the shape that reads as lock-in to the enterprise/HIPAA buyers we are selling to. iris bloqs export <id|name> [-o dir] [--attachments] [--no-markdown] [--json] Writes bloq-<id>-<slug>/ containing: - bloq.json full API payload, verbatim (fidelity copy) - items/NN-list/NNN-title.md one markdown file per item, metadata in frontmatter (the copy that stays readable without us) - files.json attachment manifest; --attachments downloads the bytes - manifest.json format version, counts, and what was NOT included Notes: - resolves by name as well as id, reusing the existing resolveBloqId - attachment fetch failures are non-fatal and counted in the manifest — a partial export beats no export, but the gap is recorded, not silent - does not stack a second H1 on bodies that already open with one (everything from `bloqs publish` carries its own title) Verified against bloq anomalyco#503: 32 lists, 350 items, 350/350 markdown files, round-tripped by both id and name. Next (not in this commit): --schedule via the existing scheduler, a workspace- level export across bloqs, and UI + API surfaces — today a non-CLI client still has no path to their own data. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

A per-bloq export answers "can I get this bloq out?". It does not answer "can I get my data out?" — you'd have to enumerate your bloqs by hand and remember to add new ones. One command, and therefore one cron line, has to capture everything. iris bloqs export --all -o ~/iris-backup Walks every bloq the user owns, exports each with the existing single-bloq logic (now factored into exportOneBloq), and writes workspace-manifest.json at the root: per-bloq counts, totals, and an explicit failures[] array. Design notes: - a per-bloq failure does NOT abort the run. One bad bloq must not cost you the entire backup, so failures are collected and reported, never silently dropped. failures[] is in the manifest whether or not it's empty. - the no-arg case now guards BEFORE the intro banner, which previously printed "Export bloq undefined" at you before saying what it wanted. Verified end-to-end against the live account: 109/109 bloqs, 778 lists, 8,924 items, 0 failures, ~3 min, 84 MB on disk. Integrity checked: 8,924 .md files on disk == 8,924 items in the manifest. Single-bloq export re-run as a regression (anomalyco#503: 32 lists / 350 items). Scheduling is deliberately NOT in this commit. A scheduled export has to run where the disk is, and the server-side scheduler cannot write to a user's machine — the honest home is a hive_task_dispatch job on the local daemon, and I'm not guessing at that payload contract. Today `--all` is cron-able as-is. Still open: UI + API surfaces — a non-CLI client has no path to their data. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Closes competitive gap G2 (bloq anomalyco#503 item #178177, bug #178193) from the Agent Orchestrator gap analysis: Hive owned the harder half (multi-machine dispatch + peer mesh) but had no cockpit, so blocked work was invisible until you thought to ask. `hive tasks` answers "what is on this node". `hive board` answers "what is blocked on me", across every node: - Aggregates the three existing sources (local bridge daemon queue, claimed pending tasks, cloud fleet history) and dedupes by task id. - Groups into NEEDS YOU / WORKING / QUEUED / DONE, mirroring AO's lane model. DONE is hidden by default — surface what's blocked, hide what's finished. - Attributes every task to a node via the node roster. - Flags queued tasks older than --stale-after (default 60m) as stuck, so silently-wedged work lands in NEEDS YOU instead of looking merely queued. - Prints an explicit degraded-source warning when a fetch fails, so a partial read can never masquerade as an idle fleet. Flags: --all --node --since --limit --stale-after --json First run on the live fleet surfaced 23 failed dispatches in 24h and 1/11 nodes online — none of it previously visible from any single command. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

…R-14 `iris imessage send` shelled out to osascript; `iris mail send` POSTed straight to the bridge; `iris run send_imessage` / `send_email` mapped onto the bridge directly. All three worked, and all three were invisible — nothing wrote lead_comms, so the log was only ever as fresh as the last time somebody remembered `atlas:comms ingest`. Measured in production (#178647): 27 of 28 leads with iMessage history were more than a week stale, several by ~2 months. The bridge is still the transport. The router is now the bookkeeper. comms-send.ts is the single call into POST /api/v1/comms/send, so there is one place that knows the wire format rather than three. `iris run` deliberately FALLS THROUGH to the bridge if the router is unreachable — it is the low-level escape hatch, and removing someone's ability to send because the API is down would be a worse failure than an unlogged send. Epic: bloq anomalyco#503 / list anomalyco#1922 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

… with the tests they went out without Gemini CLI support for `iris mcp install`: iris mcp install --client gemini -> ~/.gemini/settings.json Keyed "iris", not the canonical "IRIS OS". Gemini names every tool mcp_<server>_<tool> and then parses the server back out by splitting at the FIRST underscore, so a key containing a space or an underscore resolves to a server that does not exist and silently breaks per-server includeTools / excludeTools / trust. The entry also declares IRIS_API_KEY explicitly: Gemini force-redacts *KEY* host vars from stdio servers, and an explicit `env` is merged in after that redaction (unlike `headers`, which it expands against the SANITIZED env and would turn into an empty bearer). Verified end to end against Gemini CLI 0.55.1 — server reached, tools listed: ✓ iris: ~/.iris/bin/iris mcp serve (stdio) - Connected Also lands the tests for two fixes that were committed without them: platform-bloqs-list-filter.test.ts `bloqs items -l <list>` filtered only the single page it happened to fetch, so on bloq anomalyco#503 (563 items) it answered "No items found" for a list with 28. An empty result wearing the costume of a definitive one. Covers paging-until-satisfied and the honesty flag that separates "empty" from "I stopped looking". platform-pages-scaffold.test.ts `pages create` scaffolded a SiteFooter with no `copyright` — a prop COMPONENT_REGISTRY in the same file declares required — so it rejected the page it had just built, and since `pages push` 404s on a slug that does not exist there was no create-then-push path at all. Checks every scaffolded component against its own registry contract. 20 tests, typecheck clean. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LiYuu8tdfepPpYD5PxRHh2

`publish` reading scaffold/how-to is correct and stays — publishing from ~/.iris/how-to would push unreviewed files nobody merged. The bug was the other half: `add` wrote to ~/.iris/how-to silently, so a recipe was authored into the one directory `publish` never reads, and every surface reported success. Six recipes had been sitting there unpublishable, and nothing named them. - `add` writes to the repo by default and mirrors to ~/.iris/how-to so view/search/MCP see it without a reinstall. It prints the path it actually used. `--local` is now an explicit choice, and says plainly that the recipe cannot be published. - `how-to promote <name>|--all` — copy a local-only recipe into the repo. Bare `promote` lists what is stranded. - `list` marks local-only recipes; `publish` names them on both the dry-run and success paths. The absence was previously rendered nowhere, which is the same shape as a deploy check that cannot tell "healthy" from "not measured". Promotes four of the six stranded recipes and adds remote-support-a-family-pc. SANITISED — /how-to is public and two live recipes were leaking: pulse.md and crowdfunding-opportunities.md carried a real person's name, a real lead ID and a deal value. manage-staff-and-contracts carried a real prospect's name and email address as example values. All replaced with example data. andrew-esher-full-demo and onboard-new-client are held back deliberately — they name a prospect throughout and disclose his client pricing. The only control that caught this was reading the files by hand. A refusing pre-publish scan is specced as HOWTO-06 in bloq anomalyco#503. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GjMn3kzkiaNWMrA4REERpu

…already written `iris timeline [--since 4mo] [--source ...] [--grep ...]` — 6,330 episodes across 18 weeks, grouped by week, from four channels that already record what happened: diary 473 2026-04-24 → 08-21 GET /api/v6/diary/list (126 days already synced) git 5726 2026-04-24 → 08-21 commit subjects across all 11 nested repos opencode 79 2026-05-12 → 08-18 session `title`, 79/79 titled claude 52 2026-07-16 → 08-21 `custom-title` / `ai-title`, 53/68 titled NO MODEL, NO NEW TABLE, NO SUMMARISATION. Every line was written by a human or by the tool that produced the session, so it cannot hallucinate and costs nothing to run. Persisting these as `signal_episodes` on the corpus spine is the follow-up (TL-01, bloq anomalyco#503 list anomalyco#2124) — worth doing once the shape is proven, not before. SESSION CONTENT IS NEVER READ — titles and timestamps only. `corpus-collect.mjs` set that rule because prompts carry credentials and client data, and this inherits it rather than re-opening it. Today's session alone contains two live production passwords, which is the argument. COVERAGE IS PRINTED PER SOURCE, and that is the feature. These channels begin on different dates — Claude Code only reaches mid-July, opencode to May. A merged view without per-source ranges silently presents "nothing happened before May" as a fact about the business rather than a fact about the instrument. The first run proved the point: diary returned 0 and the coverage line said so instead of quietly dropping 473 episodes (missing user_id — the endpoint is user-scoped, and it answers for nobody without it, with a 200). Per-week truncation names what it hid ("… 616 more from git this week"), same rule as how-to. Two traps carried over from the corpus collector rather than re-learned: this monorepo has ELEVEN nested repos and NO .gitmodules, so gitlinks are read from the index; and each candidate must be its own `rev-parse --show-toplevel`, because `git -C` on an uninitialised gitlink walks UP and answers about the parent. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GjMn3kzkiaNWMrA4REERpu

productName/identifier/window title/menu/dialogs/CLI-install messaging, Cargo.toml metadata, package.json name, AppStream metainfo, and a full regenerated icon set (from the existing IRIS app icon) all move from "OpenCode" / ai.opencode.desktop to "IRIS" / io.heyiris.desktop. Updater endpoint now points at FREELABEL/iris-opencode's own release feed instead of anomalyco/opencode's, with a freshly generated signing keypair (pubkey only, committed; private key handed off out of band). Also fixes a real bug found along the way: cli.rs's CLI-install path checker was hardcoded to ~/.opencode/bin/opencode, but the actual `install` script at the repo root has installed to ~/.iris/bin/iris for a while — the desktop app's "Install CLI" / version-sync logic was checking a path the real installer never creates. Renamed the sidecar handle (opencode-cli -> iris-cli) and the desktop package's SIDECAR_BINARIES target names to iris-<platform>, matching what release.yml's `build` job actually produces. Left @opencode-ai/* package names, OPENCODE_* env vars (the desktop<->CLI subprocess protocol), and the Rust crate/lib name untouched -- internal contracts, not user-facing branding, and out of scope here. NOT done in this pass (tracked in bloq anomalyco#503 item #182113): wiring this package into release.yml at all (it isn't built by CI today), real code signing/notarization, and rewriting prepare.ts's artifact download to match release.yml's actual per-target archive shape.

Local Rust builds keep hitting ENOSPC (disk is at 99% capacity on the machine doing this work) -- adding a workflow_dispatch-only job that builds packages/desktop on a macos-latest runner and uploads the .app/.dmg as workflow artifacts, so the rebrand can be tested without local disk headroom. Unsigned, ad-hoc output for local testing only -- not wired into the real release process (release.yml is untouched). Safe to delete once real CI wiring for this package lands (tracked in bloq anomalyco#503 item #182113).

Adds build-desktop-tauri + desktop-tauri-assets, mirroring the existing build-desktop + desktop-assets pattern used for the Electron/SaaS-wrapper app: independent, continue-on-error, non-blocking, attaches to the already-created release asynchronously so a slow/failed desktop build never holds up CLI releases. Reuses the exact build steps already proven working in desktop-test-build.yml (fresh bun install -> build the iris CLI -> stage it as the Tauri sidecar -> tauri build -> zip the .app), but with the real release version instead of a hardcoded test version. Deliberately NOT done here: code signing / notarization (no Apple Developer ID cert, no Windows cert, no owned TAURI_SIGNING_PRIVATE_KEY exist for this repo -- gh secret list confirms only CROSS_REPO_PAT is set) -- updater artifact generation is explicitly disabled until a real signing key exists. Output is unsigned/ad-hoc-signed, fine for continued testing, NOT yet fine for a public landing-page download. Tracked in bloq anomalyco#503 item #182113. This wiring itself is untested end-to-end -- exercising it for real means triggering an actual tagged release (or a workflow_dispatch version bump), which publishes a real public GitHub Release. Deliberately not doing that as part of this commit.

build-desktop-tauri-windows -- same sequence proven working in desktop-test-build.yml moments ago (this is the first real Windows build of packages/desktop that has ever existed): build the iris CLI, stage it as the Tauri sidecar (bin/iris.exe -- confirmed via a live run, no ambiguity like the Mac path had), tauri build --bundles nsis, rename the installer to the IRIS-tauri-* convention so desktop-tauri-assets picks it up alongside the macOS zips. desktop-tauri-assets now needs both Tauri jobs and uploads *.exe alongside *.zip. Independent job, continue-on-error, same non-blocking pattern as everything else here. Unsigned -- no Windows code-signing cert exists for this repo. Alex's call: acceptable for now (bloq anomalyco#503 item #182113). SmartScreen will warn on first run.

fix: re-enable history shortcuts
e9634ef
Arrow keys up/down were not working for navigating through history. Looks like they were commented out in 568c047, looks like an accident, but I could be wrong.

thdxr force-pushed the dev branch from deafc7c to 780419eCompare June 29, 2025 01:57

thdxr self-assigned this Jul 2, 2025

thdxr force-pushed the dev branch from 44dce70 to c952e9aCompare July 18, 2025 17:40

jakst closed this Aug 20, 2025

mayoalexander mentioned this pull request Jun 24, 2026
feat: land remaining recent CLI features onto main (atlas-item, pathways onboard, mcp register) FREELABEL/iris-opencode#46
Merged

Navigation Menu

Uh oh!

fix: re-enable history shortcuts - #503

fix: re-enable history shortcuts#503
jakst wants to merge 1 commit into
anomalyco:devfrom
jakst:jakst/history-shortcuts

jakst commented Jun 28, 2025•
edited
Loading

Uh oh!

adamdotdevin commented Jul 1, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Uh oh!

Conversation

jakst commented Jun 28, 2025• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

adamdotdevin commented Jul 1, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

jakst commented Jun 28, 2025•
edited
Loading