v0.3.67

52 commits since v0.3.66. 104 files changed, +11090 / -3413 lines.

✨ Features

• 1df3c044 Multilanguage prompt analyzer: shared PromptAnalyzer with soft-nudge on TUI and Telegram, 6 language packs (EN/ES/FR/ID/PT/RU) (#518, #538)
• 7b7ea4d6 Plan-state copy lock: F3 plan-state copy lock and /show-plan demotion (ADR 0005)
• ad3e04ae Goal chrome: F3 goal chrome with turn-local retention (ADR 0005)
• 1ab25268 Per-heading plan prose: F3 per-heading plan prose in flow chrome (ADR 0005)
• 84b2ed5e Archive at turn settle: archive completed plan at turn settle, not mid-complete (ADR 0005)
• 9ac48392 Always-visible title: F2 always-visible title and full checklist chrome (ADR 0005)
• 356e21ba Uncollapsed flow shell: F1 uncollapsed flow shell with merged footer (ADR 0005)
• 797a6d6f Web search retry: rotate User-Agent and retry DuckDuckGo 403s (#525)
• 6c9b5084 Telegram rate-limit retry: retry telegram_send on 429 rate limits (#524)
• 14a1303d Config reload feedback: surface silent config.toml hot-reload recovery to the user (#534)
• 6ab209ce Channel context header: include chat_id/thread_id in the [Channel: ...] prompt header (#533)
• 96648fb1 Provider-aware narration cap: make folded-narration cap provider-aware (#532)
• 9eab3285 Approve/Discard flow: D3 Telegram Approve/Discard flow chrome and TUI plan overlay (#521)
• f25278c1 Dual-track prompts: D2 dual-track prompts, state-branched compaction recovery (#521)
• 93414269 Design-track commands: D1 commands, Approve validator, seed turn, design-track soft-nudge (#521)
• 756bdd05 Plan tool contract: C3 plan tool contract, design/checklist tracks, goal wiring (#520)
• f457593d Session markdown mirror: C2 session markdown mirror and plan-file-changed reloads (#520)
• 99bc0f8d Tool-loop write gate: C1b tool-loop write/bash gate and Active .md freeze (#520)
• 39641597 Lifecycle engine core: C1a lifecycle engine core, Editing/Active status model with durable pre-init flag (#520)
• 491c4d24 Merged flow message: merge pre-flow into the flow message, flow chrome sections (#519)

🔧 Fixes

• 297bcb64 task_type enum: document the full task_type enum and log the Other fallback
• 18971bea Non-empty description: require a non-empty description on every task entry point
• 675e4cf2 Checklist validation: enforce required root title and description on checklist import
• ed7476aa React-only marker: keep react-only marker so claude-cli turn ends cleanly
• 3baf00af /models provider: /models command shows session provider instead of global
• d8bd3de7 React-only turn: surface a long no-tool react-only turn instead of silently dropping it (#546)
• a1f62fe7 Goal cleanup: clear the session goal on plan discard
• 19b11029 Checklist defaults: assign order and default complexity on checklist import
• 7e1021bc Header-only cleanup: clean up the header-only flow block on a react-only turn (#544)
• b3c3eab7 Design prose display: surface design prose in /show-plan and point chrome at it
• b606e7d6 Plan keyboard re-attach: re-attach plan keyboard after flow restick
• cd178879 Bot menu commands: register plan mode commands in bot menu (#537)
• ce3c008b Shell escaping: context-aware shell escaping so double-quoted params are safe (#523)
• a327e1d9 @bot mention strip: strip @bot only as a command suffix, preserve standalone mentions (#528)
• 024e0cc2 Reaction markers: accept keyword-less <> reaction markers
• 35c5891f Working header: show the 'Working on: ' header for claude-cli turns (#527)
• 6926ba2b Reaction terminators: accept and bare > as reaction-marker terminators
• 08993632 Session model/provider: resolve Runtime Info model/provider per session at prompt injection
• cd0c4782 Attachment handling: preserve attachment names, store directed files durably, organize channel_attachments by platform (#513)
• 09ee09f0 Double gear dedup: dedup the double gear when the live header status is the bare-tool fallback

📖 Documentation

• e295b307 Seed plan-mode umbrella SSOT, cluster ADRs and specs
• 11b708f3 Correct stale status and approved_at semantics in the live spec
• 150a6572 Fix self-contradiction on when the analyzer sets pre_init_editing
• 1465fa39 Drop phantom execution_history and reflection from the spec
• 82c8e8e3 Describe async project/profile-aware session dir resolution

🧹 Miscellaneous

• 12b6386d Drop deprecated task_manager from the tool catalog
• 81496afd Drop orphaned plans/plan_tasks tables
• 54611a75 rustfmt the /execute BotCommand registration
• 31ff3990 Remove unused opencli skill pointing at unbundled binary (#536)
• 87180a31 Retire orphaned Plan/PlanTask row models, CHANGELOG for the lifecycle engine (#520)
• 89ead77f Resolve session dir async, project/profile-aware

📊 Stats

• 52 commits since v0.3.66
• 104 files changed, +11090 / -3413 lines
• 4938 tests (4938 passed, 0 failed, 29 ignored)

🗑️ Removed

• opencli built-in skill (#536): dropped the skill whose SKILL.md advertised 25+ opencli-rs dynamic tools (hn_top, twitter_trending, google_news, and others) that require a separate opencli-rs binary and a daemon on port 19825. Nothing shipped or provisioned that binary, so on a stock build the skill loaded a doc describing tools that were never registered. Usage history confirms zero successful executions of any opencli-rs tool. Removed the BUILTIN_SKILLS registration, the template directory, and the README and BRAIN_CONSTITUTION references.

✨ Features

• Plan mode UX (#521): the user-visible product on top of the lifecycle engine. /plan, /show-plan, /execute and /discard land on TUI and Telegram; the Approve validator scans the session .md (Context labels filled, at least one numbered step) and a passing /execute sets Active, stamps approved_at, and dispatches one visible seed turn that converts the numbered steps into the checklist and starts task 1. Approve and /execute are refused while a turn runs (never queued); /discard cancels the turn first. The shared soft-nudge swaps the plan hint to the design track and plan keywords durably enter Plan mode. Prompts now teach SESSION PLAN vs CHECKLIST with the locked track-selection table, every Editing turn pins the plan-mode rules, compaction recovery branches by plan state and the summary carries a harness-written PLAN STATE block, and the deprecated task_manager tool left the preamble and catalog. On Telegram, Approve/Discard inline buttons (owner-only in groups, plan: callback prefix) plus Editing/Building-checklist chrome ride the latest flow message; on the TUI, /show-plan opens a scrollable plan overlay with approve/discard keys, and the checklist strip shows Building checklist… or seed-error while the seed window is open.
• Plan lifecycle engine (#520): plans live as NoPlan / Editing / Active. The seven-value status enum collapsed to Editing and Active with a legacy map on load (Completed archives silently under agents/session/archive/, Cancelled deletes, old draft checklists stay executable), all plan IO goes through one shared store (src/utils/plan_files.rs), and a durable pre_init_editing sidecar survives restarts. A tool-loop gate enforces the Editing write policy: pre-init denies project writes but allows exploratory bash, post-init Editing denies all bash and allows writes only to the session plan .md (whose body mirrors into the JSON description on save), and Active freezes the design .md. The plan tool gained mode (design or checklist), add_tasks (with add_task kept as an alias), Active-only start/complete with deterministic Editing refusals, import rules, archive-on-last-complete, and acceptance criteria that set the session goal while a task runs; auto-approve on first start is gone, and the dormant SQLite plan path (PlanService, PlanRepository, row models) is retired.
• Telegram flow chrome merge (#519): one live flow message per turn. The pre-flow status bubble is gone; thinking and Working-on previews ride the flow header from the first activity tick, all three renderers support header-only output (empty entries still emit the header), and no-tool long turns open and settle a header-only Processing log. The ctx footer moved off final answers onto the settled flow message, plan title, checklist progress (from the live session plan JSON), and the active goal one-liner render as always-visible flow sections via one shared section builder, and crash-recovery resume now shares the same open-early, live-duration, and settled-header path as live turns.
• Shared PromptAnalyzer soft-nudge on TUI and Telegram (#518): the keyword analyzer moved from src/tui/prompt_analyzer.rs to the shared src/utils/prompt_analyzer.rs and now runs on both surfaces. Plan hints teach the live tool contract (init, add_task, start) and explicitly reject the dead create/finalize operations. Hints are LLM-only (appended to the agent input, never to Telegram display_text or TUI chat bubbles), and slash commands plus skill/user-command expansions are never analyzed on either surface.

Full Changelog: v0.3.66...v0.3.67