Stall detection now fires while agent is busy

AGENTS.md

@@ -0,0 +1,36 @@
+# AGENTS.md — amplifier-app-cli
+
+## Boundary rule: `amplifier_app_cli/data/` vs. an external bundle
+
+`amplifier_app_cli/data/` ships inside the CLI's own wheel — anything placed there loads
+for every user, every session, version-locked to the installed CLI, with no bundle
+composition step in between. That reach is exactly why it must stay small. Before adding
+anything here, run it through these three tests, in order:
+
+1. **Does it depend on something the CLI uniquely provides, that cannot move?** A slash
+   command, a settings key, a terminal affordance — something with no home outside this
+   process. If no → it belongs in an external bundle, not here.
+2. **Would a non-CLI host ever want it?** If yes → external bundle. The CLI may still
+   *include* it (compose the bundle), but must not *own* it — ownership belongs wherever
+   the capability is portable to.
+3. **Is the trigger unconditional?** If the asset is gated on settings, an env var, a flag,
+   or runtime state, the asset itself may still live here, but the compose/injection
+   *decision* stays in Python (see `runtime/config.py::_ensure_default_skills_dirs` for the
+   pattern) — never encode conditional loading in a bundle YAML that lives alongside it.
+
+If the answer to 1 is "no" or the answer to 2 is "yes," it's an external bundle question,
+not a `data/` question.
+
+**Resolution rule:** assets under `amplifier_app_cli/data/` are always resolved by
+**package-relative path** (e.g. `Path(__file__).parent.parent / "data" / "..."`), **never**
+by git URI. A git URI decouples the asset's version from the installed wheel's version —
+defeating the reason for co-locating it here in the first place. If it needs independent
+versioning, it isn't a `data/` asset.
+
+**Token budget:** this location is auto-loaded for every user, every session — its budget
+discipline is stricter than anywhere else in the ecosystem. No always-on context files
+here without an explicit, named exception recorded in this section. Prefer mechanisms
+that load on demand (skills, agent-scoped context) over anything injected unconditionally.
+
+This section exists to keep `data/` from becoming a junk drawer — re-run the three tests
+before adding, not after.

amplifier_app_cli/data/skills/goalify/PROVENANCE.md

@@ -0,0 +1,69 @@
+# goalify — rule provenance
+
+Not loaded by `load_skill`. This file exists so the lint rules in `SKILL.md`
+can be audited and re-derived. Read it when changing a rule, not when using
+the skill.
+
+## Where the rules come from
+
+Every BLOCKER traces to an observed `/goal` run that failed to terminate.
+That property is the ruleset's credibility. **Do not add a rule without a
+named run or a corpus measurement behind it.**
+
+| Rule | Origin |
+|---|---|
+| L1 — ordering/provenance constraint | A run whose condition required `PROVEN with evidence you produced yourself`. The evaluator crystallised this into a constraint on the transcript's own past and stated it could not be retroactively repaired. Stalled after 54 turns; unreachable from roughly turn 45. |
+| L2 — universal quantifier with exempt members | A run requiring evidence for `all 9 sites` in a uniform structure, where one site could not structurally produce that evidence. Stalled after 24 turns. |
+| L3 — elapsed wall-clock requirement | A run whose only available proof standard was real-world use over time. Stalled after 31 turns. |
+| L4 — human-in-the-loop dependency | A condition containing `Stop and ask me if you need a decision from me`, which contradicts unattended continuation. 12 turns consumed in 35 seconds. |
+| L5 — open enumeration | `all editing features of Publisher` — never terminated. Its sibling run in the same session, same codebase, same day (`until I have a usable app`) achieved on turn 1. |
+| L0 — cross-clause consistency | The L2 run above carried a textbook four-verdict exit vocabulary and stalled anyway, because one other sentence silently overrode it. |
+| L6 — missing disjunctive exit | Advisory, not blocking. See below. |
+
+## Why L6 is a WARNING, not a BLOCKER
+
+A lint regression over 30 scored real runs with known outcomes measured L6
+firing on 37% of all real conditions at a 9% hit rate — the largest single
+source of false positives of any rule, consistent across the tuning split,
+the held-back split, and the full set. L6 judges exit-clause presence in
+isolation and cannot see turn position or residual scope, so it routinely
+flagged short finisher-style conditions written late in a long session.
+
+On that corpus, presence of a disjunctive exit was only weakly correlated
+with actual termination — several conditions with strong exit language
+stalled anyway, for reasons L1/L2/L0 cover independently.
+
+Demoting L6 raised measured precision from 20% to 43%, recall unchanged
+at 100%. L0 and L1–L5 each fired only 1–2 times in that evaluation — too few
+observations to justify changing their classification, so none was changed.
+
+## How to read the precision number
+
+The 20%/43% figures were measured on a corpus dominated by **human-authored**
+conditions, which terminate about 96% of the time. On a population that
+rarely fails, any linter's precision is bounded by arithmetic — a rule that
+flags a condition which succeeded anyway counts as a false positive.
+
+This skill lints **agent-authored** conditions, which in the same corpus
+terminated about 60% of the time. Same rules, roughly ten times the base
+rate of true positives. **Treat 43% as a floor measured on the easiest
+available population, not as the operating precision.** Re-scoring against
+the agent-authored subset alone is the outstanding measurement.
+
+The asymmetry also justifies the operating point: a false positive costs one
+in-session rewrite pass; a false negative costs a 24-to-54-turn unrepairable
+stall. High recall at moderate precision is the correct trade here.
+
+## Standing caveats
+
+- **The corpus ages.** These rules encode `/goal` evaluator semantics as
+  observed at the time of measurement. If the goal loop's evaluator changes,
+  nothing will automatically flag that the rules have rotted.
+- **Effective sample size is smaller than it looks.** The scored runs came
+  from roughly 8 distinct sessions; runs within a session share an author, a
+  project, and phrasing habits.
+- **The skill drafts; the human edits.** Automating goal authoring is itself
+  the thing that raises failure rates (60% vs 96%). The lint is the bet that
+  it closes that gap, and that bet has not been measured end to end. The
+  human review step is load-bearing, which is why the skill offers and never
+  auto-runs.

amplifier_app_cli/data/skills/goalify/SKILL.md

@@ -0,0 +1,188 @@
+---
+name: goalify
+description: >
+  Compose a /goal stop-condition from the current conversation and lint it
+  against known termination-failure patterns before showing it to the user.
+  Use when the user wants to turn the current task into a /goal loop, asks to
+  "goalify this", wants a stop condition for autonomous work, says "write a
+  goal condition", "make this a /goal", "turn this into a goal", or asks for
+  help wording a condition for /goal.
+user-invocable: true
+version: 1.1.0
+license: MIT
+---
+
+Run this procedure yourself, in the current conversation. Do not delegate it
+to a sub-agent or forked session — Phase 1 reads the live transcript.
+
+$ARGUMENTS
+
+If the user supplied focus text above, use it to scope Phase 1. If empty,
+extract from conversation alone — do not ask the user to restate what they
+already said.
+
+---
+
+## Phase 1 — Extract
+
+From the conversation so far, determine:
+
+- **The target end state.** What does "finished" concretely look like? It
+  must be a state that can be checked, not an activity that can be performed
+  indefinitely. ("a usable app" is checkable; "finish building the project"
+  is not — there is no test for "finished building".)
+- **What is already done.** Re-read the transcript for completed sub-tasks,
+  passing tests, merged changes, or resolved questions. These become
+  candidates for the KNOWN section (Phase 2) or for narrowing scope
+  (SCOPE-OUTS).
+- **What is explicitly NOT required.** Anything the user has ruled out,
+  deferred, or said isn't needed. This is the raw material for SCOPE-OUTS.
+
+If the end state genuinely cannot be determined from context (not merely
+effortful to determine), ask one direct question. Otherwise proceed —
+guessing and then showing your extraction for correction is faster than
+front-loading a question the transcript already answers.
+
+## Phase 2 — Compose
+
+Emit a candidate condition using this structure. Every element is required
+unless marked optional.
+
+1. **One-sentence outcome** naming a checkable end state (not an activity).
+2. **Disjunctive exit**: the condition must be satisfiable by *either*
+   reaching the end state *or* conclusively demonstrating it cannot be
+   reached (naming the blocker). Never phrase a condition with only one exit.
+3. **Per-item negative terminal**, if the condition lists multiple items
+   (tasks, sites, phases, experiments). Each item must be able to resolve to
+   its own PASS / FAIL / BLOCKED-with-named-reason — a blocker on one item
+   converts *that item* to a residual; it must not block the whole goal.
+4. **SCOPE-OUTS section** — an explicit list of what is *not* required. Write
+   this by directly converting anything from Phase 1's "not required" list
+   into a plain negative statement (e.g. "No production soak time required."
+   / "Uniformity across all N items is NOT the goal.").
+5. **KNOWN section (optional)** — facts already established, so the actor
+   doesn't re-derive them. Label it explicitly as a speed aid: it prevents
+   wasted turns, it does not by itself prevent stalls, so it never replaces
+   items 1–4.
+
+## Phase 3 — Lint
+
+Check the full composed document against every rule below. Work through all
+BLOCKERS first; a document with any BLOCKER triggered is not ready to show.
+Then check WARNINGS, which are advisory and do not block presentation.
+
+**Read the whole document for each check.** Several of these rules are only
+detectable by considering the document as one system — a single clause
+elsewhere can silently defeat a correct-looking rule everywhere else. Do not
+scan for keywords in isolation and stop at the first clean-looking match.
+
+### BLOCKERS — fix all before presenting
+
+- **L1 — Ordering/provenance constraint on the transcript's own history.**
+  Any phrasing that requires evidence to precede, or be produced independent
+  of, events that already exist in the transcript (e.g. "verify it yourself,
+  then state what you verified", "proof must precede the claim", "evidence
+  you produced yourself" applied to something already reported by a
+  sub-agent or prior turn). This class of requirement is **unrepairable** —
+  no later turn can change what already happened earlier in the transcript,
+  so it cannot be fixed by adding more work. If the condition constrains
+  ordering, it must constrain only *future* actions, never re-litigate what
+  is already in the history.
+
+- **L2 — Universal quantifier over a set with possibly-exempt members.**
+  "all N", "every X", "each of the Y", "uniform/uniformity", "complete
+  parity", applied to a set, is a blocker **unless** each item individually
+  carries a negative terminal (see Compose #3) or the condition names which
+  members are exempt and why. Without one of those, a single member that
+  cannot structurally produce the required evidence makes the whole
+  condition permanently unsatisfiable.
+
+- **L3 — Elapsed wall-clock requirement.** Anything that requires real time
+  to pass beyond the current session: "production soak", "after N days of
+  use", "monitor over time", "verify in real-world use". A single session
+  cannot advance wall-clock time; this can never be satisfied in-session.
+
+- **L4 — Human-in-the-loop or external-actor dependency mid-loop.**
+  "stop and ask me if you need a decision", "once a reviewer merges this",
+  "wait for approval before continuing". This directly conflicts with
+  unattended continuation — the loop will halt waiting on an event that a
+  condition-checking loop cannot itself produce.
+
+- **L5 — Open enumeration.** Scope phrased as an unbounded or unenumerated
+  set: "all editing features of X", "complete parity with Y", "everything
+  needed to fully support Z". An evaluator can always name one more item
+  under this phrasing, so it never terminates. Convert to a closed, named
+  list, or to a single representative artifact.
+
+- **L0 — Cross-clause consistency (meta-rule).** *An escape hatch is only as
+  strong as the strictest other clause in the same document.* After
+  confirming L1–L5 pass individually and a disjunctive exit exists, re-read
+  the document once more asking only: **is there any other sentence, anywhere
+  in the document, that is stricter than the stated exit and would override
+  it?** A document can have a textbook-perfect exit clause and still be
+  unsatisfiable because one unrelated sentence elsewhere re-imposes an L1–L5
+  style constraint the exit clause doesn't cover. Confirming an exit clause
+  exists is not sufficient — confirm nothing else in the document is
+  stricter than it.
+
+### WARNINGS — advisory, do not block presentation
+
+- **L6 — Missing disjunctive exit.** The document should state achievement
+  *or* a way to conclusively end in "not achievable, here is why" (see
+  Compose #2). Flag and fix its absence where practical, but do not block
+  presentation on it alone.
+- **W1** — Multiple items are listed but not all of them carry their own
+  negative terminal (some do, some don't).
+- **W2** — No clause asking the actor to show evidence inline in the
+  transcript as it's produced, rather than only asserting a result.
+- **W3** — Scope reads like more than one session's worth of work (multi-week
+  rollout language, coordination across many independent repos/teams,
+  phased production deployment).
+- **W4** — The condition contains a cautionary anecdote or narrative about a
+  failure mode (e.g. "don't repeat what went wrong last time", "make sure
+  this doesn't stall like before") rather than a plain instruction. Any such
+  narrative addressed to the actor is read by the evaluator too, and can
+  silently become a criterion the evaluator judges against instead of
+  guidance the actor merely follows. State requirements as plain criteria,
+  never as stories.
+
+  **This applies to the condition you are composing right now.** Write every
+  clause as a direct instruction to the actor, never as a story about a past
+  run. If you catch yourself writing "so that we don't repeat X", rewrite it
+  as the direct requirement it implies, with no reference to the incident.
+
+### If a BLOCKER cannot be cleared
+
+Rewrite and re-check. Allow up to three rewrite passes. If a BLOCKER still
+fires after three passes, stop and surface the specific tension to the user
+by name (e.g. "the user's own request requires enumerating an open-ended set
+— L5 fires no matter how I phrase it; how would you like to bound this?").
+Do not present a condition that still fails a BLOCKER.
+
+---
+
+## Output format
+
+Always output the condition inside a fenced code block — terminal reflow
+will otherwise destroy its multi-line structure. Follow it with the lint
+report as a table, then offer (do not auto-run) `/goal`.
+
+```
+<the condition text>
+```
+
+| Rule | Result | Note |
+|------|--------|------|
+| L0 | no known pattern detected | ... |
+| L1 | no known pattern detected | ... |
+| L2 | no known pattern detected | ... |
+| L3 | no known pattern detected | ... |
+| L4 | no known pattern detected | ... |
+| L5 | no known pattern detected | ... |
+| L6, W1–W4 | (list only the ones that fired) | ... |
+
+A clean table means no *known* failure pattern was detected — not that the
+condition is validated. Say so if the user reads it as a guarantee.
+
+Then: "Pass this to `/goal` to start the loop — want me to run it now, or
+would you like to adjust anything first?"

amplifier_app_cli/data/skills/goalify/examples/known-bad-L1-ordering.md

@@ -0,0 +1,25 @@
+# Known-bad example — L1 (ordering/provenance constraint)
+
+Used to sanity-check the lint in `SKILL.md` Phase 3. This condition is
+modeled on a real run that stalled for many turns because the ordering
+constraint it imposes on the transcript's own history became impossible to
+satisfy once evidence had already been reported by a sub-agent.
+
+## Condition text
+
+```
+Done when the migration is PROVEN complete, with evidence you produced
+yourself. A sub-agent's report is not proof. Verify it yourself, then state
+what you verified — or show a named blocker and stop.
+```
+
+## Expected lint result
+
+- **L1: FAIL** — "with evidence you produced yourself" combined with "verify
+  it yourself, then state what you verified" imposes an ordering constraint
+  on the transcript's own history. If a sub-agent already reported a result
+  earlier in the transcript, no later turn can retroactively make that
+  report "verified by you first" — the requirement is unrepairable once the
+  transcript already contains the sub-agent's report.
+- L6 present (disjunctive exit exists: "or show a named blocker and stop"),
+  but L1 alone should be sufficient to fail this condition.

amplifier_app_cli/data/skills/goalify/examples/known-bad-L2-quantifier.md

@@ -0,0 +1,30 @@
+# Known-bad example — L2 / L0 (universal quantifier + cross-clause override)
+
+Used to sanity-check the lint in `SKILL.md` Phase 3. Modeled on a real run
+that stalled because a uniform requirement applied to every member of a set
+made the condition permanently unsatisfiable once one member could not
+structurally produce the required evidence.
+
+## Condition text
+
+```
+Done when all 9 sites are migrated in a uniform structure, verified working,
+or proven impossible with a named blocker.
+```
+
+## Expected lint result
+
+- **L2: FAIL** — "all 9 sites" + "uniform structure" is a universal
+  quantifier over a set, with no per-item negative terminal and no named
+  exemption for a site that cannot physically satisfy "uniform structure."
+- **L0: FAIL** — the document has a disjunctive-looking exit ("or proven
+  impossible with a named blocker"), but that exit applies to the *goal as a
+  whole*, not per-site. The stricter clause ("uniform structure" across all
+  9) is not covered by the escape hatch, because the escape hatch only
+  fires once — it can't let 8 sites succeed uniformly while 1 is blocked and
+  still call the set "uniform." Confirming the exit clause exists is not
+  enough; it does not cover the stricter clause.
+- L6 nominally present (a disjunctive-shaped phrase exists) but does not
+  actually rescue the condition — this is exactly the L0 case: an escape
+  hatch that reads as satisfied but is overridden by a stricter clause
+  elsewhere in the same document.