@event4u/agent-config 1.13.0 → 1.15.0

package/docs/contracts/STABILITY.md

@@ -0,0 +1,95 @@
+---
+stability: stable
+---
+
+# Stability policy for `docs/contracts/`
+
+This directory ships the **public contract surface** of `agent-config`.
+Every file here declares a stability level in YAML frontmatter:
+
+```yaml
+---
+stability: stable | beta | experimental
+---
+```
+
+The level dictates how the contract may be linked from the public
+surface (README, AGENTS.md, `docs/architecture.md`) and what kind of
+change to it requires what kind of release.
+
+## Levels
+
+### `stable`
+
+- **Breaking change** requires a **SemVer-major** bump (`X.0.0`).
+- README, AGENTS.md, and `docs/architecture.md` MAY link to it
+  without a marker.
+- Typical content: settled ADRs (decisions don't reopen without a
+  successor ADR); fully released contracts that have shipped through
+  one major release without breaking.
+
+### `beta`
+
+- **Breaking change** is allowed in a **minor-version** release
+  (`1.X.0`), provided the change appears in `CHANGELOG.md` under a
+  `### Breaking` heading.
+- README, AGENTS.md, and `docs/architecture.md` MAY link to it,
+  **provided** the link text or the surrounding sentence carries a
+  visible `(beta)` marker.
+- Typical content: flow contracts and recipes that are shipped and
+  load-bearing, but whose surface is expected to evolve before a
+  SemVer-major lock.
+
+### `experimental`
+
+- **Breaking change** is allowed in **any release** (including
+  patches), with a CHANGELOG note.
+- README, AGENTS.md, and `docs/architecture.md` MUST NOT link to
+  experimental contracts. Only the index inside `docs/contracts/`
+  may reference them.
+- Typical content: spike artefacts, runtime modules in pilot status,
+  early API drafts not yet wired into a roadmap-locked phase.
+
+## Frontmatter requirement
+
+Every `*.md` under `docs/contracts/` (except this `STABILITY.md` file
+itself) MUST start with a YAML frontmatter block declaring `stability:`.
+
+The link checker (`scripts/check_public_links.py`, P0.1b) reads this
+frontmatter and:
+
+- **fails CI** when README / AGENTS.md / `docs/architecture.md` links to
+  a contract marked `experimental`, to a missing target, or into
+  `agents/contexts/` (internal surface).
+- **warns** (non-fatal in default mode; fatal under `--strict`) when a
+  public-surface link to a `beta` contract has no `(beta)` marker in
+  the surrounding text.
+- ignores `stable` links.
+
+Run `task check-public-links` locally; `task ci` invokes the same
+checker in default mode.
+
+## Promotion path
+
+`experimental → beta → stable`. Demotion is allowed (e.g. `stable →
+beta` to permit a refactor) but appears in `CHANGELOG.md` under
+`### Breaking` and gets a SemVer-major bump.
+
+Promotion criteria:
+
+- `experimental → beta` — at least one shipped roadmap phase has
+  consumed the contract end-to-end without a breaking change.
+- `beta → stable` — at least one SemVer-minor release has shipped
+  with the contract unchanged, or the contract has been explicitly
+  frozen as part of a roadmap step.
+
+## Current contracts
+
+See the file headers themselves for current levels. The frontmatter is
+the authoritative source — this list is illustrative, not load-bearing,
+and is generated by `scripts/check_public_links.py --list`.
+
+## See also
+
+- [`agents/roadmaps/road-to-post-pr29-optimize.md`](../../agents/roadmaps/road-to-post-pr29-optimize.md) — P0.1, P0.1a, P0.1b
+- [`docs/architecture.md`](../architecture.md) — package architecture overview

package/docs/contracts/adr-chat-history-split.md

@@ -0,0 +1,132 @@
+---
+stability: beta
+---
+
+# ADR — Chat-history rule split
+
+> **Status:** Decided · 2026-05-02 · scoped under
+> [`road-to-post-pr29-optimize.md`](../../agents/roadmaps/archive/road-to-post-pr29-optimize.md)
+> Phase 2 item P2.1.
+> **Context:** AI #1, AI #3, AI #5 review of PR #29 flagged the
+> 200-line monolithic `rules/chat-history.md` as the rule the agent
+> revisited 12+ times during the 1.14.0 cycle — three independent
+> concerns coupled into one Iron Law block.
+> **Builds on:** [`adr-implement-ticket-runtime.md`](adr-implement-ticket-runtime.md)
+> — engine path was added to the existing rule mid-flight, which
+> exposed the coupling.
+> **Defers to:** Phase 2 P2.9 (memory-visibility surface) for the
+> reuse of the heartbeat plumbing.
+
+## Decision
+
+`rules/chat-history.md` is split into **three sibling `always` rules**,
+each owning one orthogonal concern and a clear single Iron Law:
+
+| Rule | Iron Law owned | What it answers |
+|---|---|---|
+| `chat-history-ownership.md` | Foreign/Returning/Missing handshake — whose file is this? | Activation, two-paths table (HOOK/ENGINE/CHECKPOINT/MANUAL), Foreign/Returning prompts |
+| `chat-history-cadence.md` | turn-check + append gates — when to write? | Turn-start gate, append cadence (`per_turn`/`per_phase`/`per_tool`), `OWNERSHIP_REFUSED` handling |
+| `chat-history-visibility.md` | Heartbeat marker — how does the user see drift? | Heartbeat invocation, `on`/`off`/`hybrid` modes, memory-typing-the-marker hard-rule |
+
+The original `rules/chat-history.md` is deleted. Cross-references
+across the three splits are explicit one-line backlinks at the top of
+each file. The CHECKPOINT-path Iron Law (the three-gate block) is
+re-anchored: ownership owns gate 1 (turn-check), cadence owns gate 2
+(append), visibility owns gate 3 (heartbeat).
+
+## Why this was a real question
+
+Three options were on the table:
+
+1. **Leave the monolith, add sub-section index.** Rejected: the
+   12-iteration churn was on individual sub-sections (heartbeat
+   visibility modes, foreign/returning prompts, cadence frequencies)
+   touching unrelated parts of the file. Index doesn't separate
+   concerns; it only renames the problem.
+2. **Move the Iron Law block to a context doc, leave the rule as a
+   thin pointer.** Rejected: contexts are reference, rules are
+   triggers — the three gates ARE the trigger surface. Demoting them
+   to context bleeds enforcement.
+3. **Three sibling rules, one concern each (chosen).** Adopted
+   because each concern has a distinct lifecycle: ownership is a
+   first-turn handshake (rare event, high stakes), cadence is per-
+   phase (high frequency, low stakes per write), visibility is per-
+   reply (every turn, observability-only). Coupling them produced the
+   "memory-typing the marker" failure mode we hardened against in
+   1.14.0 — separating them makes future hardening surgical.
+
+## Sizing & always-rule budget
+
+Pre-split: `chat-history.md` = 201 lines. Always-rule budget pre-split = 2207 lines across 18 rules.
+
+Post-split target (frontmatter + cloud-behavior + interactions
+sections triplicated):
+
+| File | Target lines |
+|---|---:|
+| `chat-history-ownership.md` | ≤ 95 |
+| `chat-history-cadence.md` | ≤ 80 |
+| `chat-history-visibility.md` | ≤ 65 |
+| **Total** | **≤ 240** |
+
+Net always-rule budget delta: **+39 lines** (≈ 1.8 % of the current
+2207-line total). Within the ~49k-token target ceiling tracked in
+`road-to-governance-cleanup.md`.
+
+## Consequences
+
+**Wins**
+
+- Each rule has one Iron Law instead of three. Future edits to
+  cadence don't risk touching ownership prompt mechanics.
+- The `chat-history-visibility.md` rule becomes a clean reuse target
+  for Phase 2 P2.9 (memory-visibility surface).
+- Refactoring the engine-side hooks (`chat_history_turn_check.py`,
+  `chat_history_append.py`, `chat_history_heartbeat.py`) is now
+  1:1 with rule files — one rule per hook.
+
+**Costs / migration surface**
+
+- Every cross-reference to `chat-history` across the package must be
+  updated to point at the right split. Inventory: ~30 files (rules,
+  templates, scripts, commands, contexts, README, AGENTS.md).
+- `docs-sync.md` table needs the new triple instead of the singleton.
+- `scripts/check_references.py` runs catch broken links; CI gates the
+  split.
+- Any consumer project carrying a project-level override of
+  `chat-history` must also split (or alias) — handled by an explicit
+  migration note in `docs/migrations/commands-1.15.0.md` (extended
+  for this rule split).
+
+**Reversibility**
+
+Split is a structural refactor, not a behavior change. Reverting
+would mean re-concatenating the three files — mechanical but loud
+in CI. The CHECKPOINT-path Iron Law numbering is preserved, so
+agents that learned "gates 1-2-3" still find the same numbered
+sequence, just across three files.
+
+## Out of scope
+
+- Behavior changes to `scripts/chat_history.py` — the script's
+  CLI surface is unchanged.
+- Adding new cadence frequencies or visibility modes — separate
+  decision, not part of this split.
+- Removing the CHECKPOINT-path entirely — Cline-on-Windows still
+  needs cooperative gates; that's a Phase 3+ conversation.
+
+## Implementation plan (this PR)
+
+1. Create the three new rules under `.agent-src.uncompressed/rules/`.
+2. Delete `.agent-src.uncompressed/rules/chat-history.md`.
+3. Run `bash scripts/compress.sh --sync` to project the change into
+   `.agent-src/`, `.augment/`, `.claude/`, `.cursor/`, `.clinerules/`,
+   `.windsurfrules`.
+4. Update every cross-reference (~30 files) — `check-refs` is the
+   gate.
+5. Append a migration note to `docs/migrations/commands-1.15.0.md`
+   under a new "Rule splits" section.
+6. Regenerate `agents/roadmaps-progress.md`.
+7. Verify: `task ci` (lint-skills, check-refs, check-public-links,
+   check-compression, counts-check, lint-readme, runtime-e2e,
+   roadmap-progress-check, lint-readme).

package/docs/contracts/adr-command-suggestion.md

@@ -0,0 +1,146 @@
+---
+stability: stable
+---
+
+# ADR — Context-Aware Command Suggestion
+
+> **Status:** Decided · Phases 1–7 shipped · 2026-04-30
+> **Roadmap:** [`road-to-context-aware-command-suggestion.md`](../../agents/roadmaps/road-to-context-aware-command-suggestion.md)
+> **Rule:** [`command-suggestion`](../../.agent-src.uncompressed/rules/command-suggestion.md)
+> **Eligibility table:** [`command-suggestion-eligibility.md`](command-suggestion-eligibility.md)
+> **Engine:** `scripts/command_suggester/`
+> **Orthogonal to:** R1 (`adr-work-engine-rename.md`) — the suggester is a
+> read-only shortcut *finder*; it never invokes a command. Engine boundaries
+> stay intact.
+
+## Decision
+
+Add a **deterministic, read-only** suggestion layer that scans every
+non-`/`-prefixed user turn against eligible command frontmatter
+(`suggestion.trigger_description` + `trigger_context`) and surfaces matches
+as a **numbered-options block** with the user-interaction Iron Law shape:
+
+- N command options, neutral labels (no inline `(recommended)` tag).
+- Always-last "Just run the prompt as-is, no command" escape hatch.
+- Exactly one `**Recommendation: K — /command**` line, omitted when the
+  top two scores are within `0.05` of each other.
+
+The suggester **never executes** anything — option-N is what triggers the
+standard slash-command flow on the next turn (with all per-command halts
+intact). The user picks every time.
+
+## Why this was a real question
+
+A maintainer who types "Setze Ticket ABC-123 um" got a generic
+implementation walk-through, not the `/implement-ticket` engine. The
+package owns 75 commands; the agent has the list in context but no
+contract telling it "when the prompt looks like *this*, offer *that*
+command". Three forces converged:
+
+1. **Discoverability gap.** Users learn `/command` only when documented;
+   most never type it.
+2. **Safety anchor.** Auto-execution is non-negotiably out (`scope-control`,
+   per-command halts). Anything we ship cannot weaken that floor.
+3. **Determinism.** ML-based intent detection is non-replayable; goldens
+   would drift. Pattern + keyword matching keeps GT-CS replays byte-stable.
+
+Three alternatives were rejected:
+
+- **Auto-route to commands** (bypass the user). Rejected: violates the
+  "user always picks" anchor; one wrong route inside `/implement-ticket`
+  is a multi-hour blast radius.
+- **Trigger-based command auto-prefix** (silently rewrite the prompt
+  to `/command <prompt>`). Rejected: same anchor breach, plus the user
+  loses the "skip" option.
+- **LLM intent classifier**. Rejected: non-deterministic, untestable in
+  the GT-CS harness, and the eligibility table is small enough that
+  pattern matching reaches the same precision at zero cost.
+
+## Eligibility rubric
+
+A command is **eligible** (default `true`) unless:
+
+- Invocation is **intentional-only** (`/onboard`, `/package-reset`,
+  `/mode`, `/agent-handoff`, `/chat-history-clear`).
+- Triggers would overlap so heavily with normal prose that the floor +
+  cooldown can't suppress the noise.
+- The command has no natural-language signal distinct from neighbours.
+
+The locked table is
+[`command-suggestion-eligibility.md`](command-suggestion-eligibility.md).
+Eligibility changes are roadmap follow-ups, not drive-by edits.
+
+## Anti-noise heuristics
+
+`rank.py` drops matches that fail any of:
+
+- Score below `confidence_floor` (default `0.6`; per-command override
+  via frontmatter).
+- Single match within `floor + 0.1` of the floor and no structural
+  bonus (high uncertainty isn't worth interrupting).
+- Short prompt (< 6 words) hitting > 2 commands without a structural
+  bonus (too vague to disambiguate).
+- Pure continuation phrase (`ok`, `weiter`, `mach weiter`, `go on`,
+  `continue`) with no structural bonus.
+- Same `(command, evidence)` pair shown within `cooldown_seconds`
+  (default `600`, per-command override).
+
+Structural bonuses (ticket key, file path, branch shape) override
+suppression when they signal real intent.
+
+## Hardening — what suggestion must never do
+
+The rule (`.agent-src.uncompressed/rules/command-suggestion.md`)
+binds the engine to five non-negotiables, mirrored as goldens:
+
+1. **No execution without user pick.** `SUGGEST. NEVER INVOKE.` is the
+   first Iron Law in the rule.
+2. **No multi-question stacks.** Clarification (`ask-when-uncertain`)
+   wins; suggestion stays silent that turn.
+3. **No conversation hijack.** `enabled: false`, no matches above floor,
+   or a senior rule active → zero output.
+4. **No echo-trigger.** `sanitize.py` strips fenced/inline code and the
+   engine's own previous block shape (`> N. /command — …`,
+   `**Recommendation: N — …**`) before scoring.
+5. **Hard subordination.** `scope-control`, `ask-when-uncertain`,
+   `verify-before-complete`, and any active role-mode contract outrank
+   suggestion.
+
+GT-CS1 through GT-CS9 (`tests/test_command_suggester_goldens.py`) lock
+the contract end-to-end.
+
+## Three opt-out paths
+
+| Path | Mechanism | Scope |
+|---|---|---|
+| Global | `commands.suggestion.enabled: false` in `.agent-settings.yml` | Whole project |
+| Per-command | `commands.suggestion.blocklist: [/cmd]` | Specific command stays as-is |
+| Per-conversation | `/command-suggestion-off` directive | Until user re-enables or chat ends |
+
+All three are deterministic and tested. `/command-suggestion-on`
+re-enables mid-conversation.
+
+## Consequences
+
+**Positive.** The maintainer's intent surfaces commands they didn't
+remember to type. Anti-noise heuristics keep the layer near-silent on
+prose. Goldens replay byte-stable; GT-CS9 covers adversarial echo.
+
+**Negative.** A new always-on rule adds context budget — kept under
+size budget by the size-enforcement rule. Per-command frontmatter now
+carries `suggestion.*` keys; the linter enforces them so drift is
+caught early.
+
+**Open.** Trigger drift over time — a command's
+`trigger_description` gets stale as conventions move. Mitigation: the
+artefact-engagement telemetry pipeline (default-off) measures
+suggestion pick-rate per command, surfacing weak triggers as
+retirement candidates without a hard SLA.
+
+## See also
+
+- [`command-suggestion`](../../.agent-src.uncompressed/rules/command-suggestion.md) — runtime rule
+- [`command-suggestion-eligibility.md`](command-suggestion-eligibility.md) — locked eligibility table
+- [`road-to-context-aware-command-suggestion.md`](../../agents/roadmaps/road-to-context-aware-command-suggestion.md) — phased delivery
+- [`adr-prompt-driven-execution.md`](adr-prompt-driven-execution.md) — `/work` entrypoint that explicit slash invocations route to
+- [`agent-settings`](../../.agent-src.uncompressed/templates/agent-settings.md) — `commands.suggestion.*` reference

package/docs/contracts/adr-implement-ticket-runtime.md

@@ -0,0 +1,122 @@
+---
+stability: stable
+---
+
+# ADR — `/implement-ticket` runtime: Python
+
+> **Status:** Decided · Phase 0 spike closed · 2026-04-22
+> **Context:** [`implement-ticket-flow.md`](implement-ticket-flow.md) ·
+> [`road-to-implement-ticket.md`](../../agents/roadmaps/road-to-implement-ticket.md)
+> **Supersedes:** the `Runtime: TBD` placeholder in `implement-ticket-flow.md`.
+
+## Decision
+
+The `/implement-ticket` orchestrator ships on **Python 3.10+** (stdlib + `pyyaml`).
+
+Bash is rejected as the runtime. It stays only where it already lives —
+the install driver, compression helper, and test runner — not the
+delivery flow.
+
+## Why this was a real question
+
+The repo already ships a Bash install + compression toolchain. A
+shell-native dispatcher would have reused the existing muscle memory and
+avoided adding Python as a delivery-runtime dependency. The spike
+existed to verify whether Bash could still carry the 8-step linear flow
+once real orchestration needs showed up (structured state, blocked/partial
+outcomes, metrics, tests). Result: not well.
+
+## Spike scope
+
+Both prototypes implement the same shape from
+`implement-ticket-flow.md`:
+
+- 8 linear steps (refine → memory → analyze → plan → implement → test →
+  verify → report)
+- `DeliveryState` as the only thing shared between steps
+- Three terminal outcomes: `success`, `blocked`, `partial`
+- Metrics JSON line per run (Q38 decision)
+- Identical fixtures: one clean ticket, one ambiguous ticket that must
+  block at `refine` with three numbered questions
+
+Sources: Phase-0 spike (`spike/implement-ticket/{bash,python}/`) — deleted after Phase 1 shipped; the evidence inlined below is the only surviving record.
+
+## Evidence (measured, not asserted)
+
+### Wall-clock mean (20 runs, 3 warmup, macOS ARM)
+
+| Scenario | Bash | Python | Winner |
+|---|---|---|---|
+| Clean, all 8 steps | 104 ms | 36 ms | **Python 2.9×** |
+| Blocked at step 1  |  60 ms | 35 ms | **Python 1.7×** |
+
+Harness was `spike/implement-ticket/bench.sh` (20 runs, 3 warmup, macOS ARM);
+raw numbers preserved in this ADR after the spike directory was deleted.
+
+Bash cost is dominated by per-step fork/exec (`yq`, `jq`, `perl`, step
+script). Every step added extends the Bash gap. Python keeps the whole
+dispatch in-process; step count is effectively free.
+
+### Test-writability
+
+| | pytest (Python) | shell asserts (Bash) |
+|---|---|---|
+| Lines to cover 3 scenarios | 59 | 72 |
+| Assert target | typed `state.outcomes`, `state.questions` | stdout string-contains |
+| Suite runtime | 0.01 s | ~2.5 s (15 subprocess launches) |
+| Test isolation | in-process `dispatch(state)` | fork per assertion |
+| IDE affordances | autocomplete, types, stack traces | grep only |
+
+### Error-propagation ergonomics
+
+- **Python:** outcomes flow through typed `StepResult`/`Outcome`
+  enums. One `dispatch()` return tuple covers success/blocked/partial.
+  Zero string parsing.
+- **Bash:** outcomes are encoded in exit codes (0/10/20), propagated
+  through `set +e` dances per step, mirrored into a JSON file with
+  `jq --arg` round-trips. Every step handler re-reads/rewrites the
+  state file. Brittle.
+
+### Source size
+
+Roughly equal (Bash 203 / Python 207 lines), but the Bash total is spread
+across 9 files with per-file shebang + `set -euo pipefail` boilerplate;
+Python is 3 files with shared types.
+
+## Tradeoffs we accept
+
+- **New hard dependency on Python 3.10+ and `pyyaml`.** Mitigated: Python 3
+  is already a build/test dependency of this repo (linters, compression,
+  `update_counts.py`). `pyyaml` is already pinned in `pyproject.toml` /
+  `requirements-*.txt`. Zero new install surface for contributors.
+- **We lose the "just-shell" story.** The install script stays Bash. The
+  delivery runtime being Python is a deliberate split: install is
+  side-effecting system work, delivery is structured data transformation.
+- **Python per-invocation boot is ~35 ms.** Accepted — it's flat, not
+  proportional to step count, and well below the user-perception floor.
+
+## Non-goals of this decision
+
+- Does not bless Python for every future spike — each flow decides
+  on its own evidence.
+- Does not commit to a specific framework (click, typer, bare argparse);
+  that is chosen during Phase 1 and kept minimal.
+- Does not move the compression/install scripts off Bash.
+
+## Consequences — unblocks
+
+- Phase 1 of `road-to-implement-ticket.md` can start: wire real step
+  handlers onto the dispatcher shape spiked here.
+- `implement-ticket-flow.md` will have its `Runtime: TBD` line replaced
+  with `Runtime: Python 3.10+` when Phase 1 lands.
+- Metrics contract (Q38, JSON lines under `agents/logs/implement-ticket/`)
+  is already demonstrated by both prototypes.
+
+## Follow-ups (not part of this ADR)
+
+- If later optimisation is needed, port the numbers above into a
+  `task bench-implement-ticket` target against the real handlers — the
+  original `bench.sh` lives only in this repo's git history now (see
+  commit `79f30e7`).
+- Decide CLI framework during Phase 1 (defer — argparse is enough for
+  the orchestrator skeleton).