agentleFS
JetXu-LLMJetXu-LLM/DocMason/skills/canonical/ask/SKILL.md

ask

Skill152 starsChanged 4 months agoApache-2.0View on GitHub

Accept an ordinary user question inside a DocMason workspace, route it to the right inner workflow, and preserve conversation-native logs automatically.

72 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use daily-arxiv, exp-eval and 10 more.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • LambenthanSkill

    daily-arxiv

    每日从 arXiv 拉取新论文,过滤相关性,auto-ingest 高优先级论文,检测 SOTA 更新

    Also in 5 projects that use this

  • LambenthanSkill

    exp-eval

    实验判决门:Review LLM 独立评判实验结果 → 4 种判决路径 → 自动更新 claims confidence、ideas status、graph edges

    Also in 5 projects that use this

  • skyllwtSkill

    exp-pilot-eval

    Pilot result evaluation — read pilot results, apply success criteria, update idea page (pilot_result, failure_reason if failed), generate PILOT_VERDICT_REPORT. Called by /ideate Phase 5 after /exp-pilot-run.

    Also in 4 projects that use this

  • skyllwtSkill

    exp-pilot-run

    Pilot experiment execution — read Pilot Spec YAML, write pilot code, run experiment(Confirm with the user before operation and require the applicant to conduct manual inspection), return results. Called by /ideate Phase 5. Does NOT modify wiki pages or judge pass/fail.

    Also in 4 projects that use this

  • phanghonghaoSkill

    poster

    HTML/CSS A4 海报生成。从沙龙/活动 Markdown 文件生成科技风 A4 海报 PDF/PNG。Use when user mentions: poster, 海报, 海报生成, make poster, generate poster, 活动海报, 沙龙海报.

    Also in 4 projects that use this

  • +6 more from Lambenthan
  • +1 more from skyllwt

The file

Show its rules and text

We split this skill into the rules it gives an agent

62 rules. 0 are common, 21 are distinctive, and 41 are very specific to this project.

  • 0

    Common

    Most of the 60 edited versions we read keep these.

  • 21

    Distinctive

    This author's own choices. Worth a look.

  • 41

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Use this skill as the default top-level workflow for a new ordinary user request in this workspace unless the request is clearly explicit …

    line 10
  2. Very specific:

    If the request is clearly an explicit setup, status, sync, adapter-maintenance, or operator-review request, switch to the matching …

    line 15
  3. Very specific:

    prefer the real native CODEX_THREAD_ID identity from the execution environment.

    line 83
  4. Very specific:

    do not read other workflow skills, workspace-status, workspace-bootstrap, source search, implementation source, or tests first just to …

    line 84
  5. Very specific:

    use those surfaces only after open returns a governed blocker, waiting state, or explicit operator route

    line 85
  6. Very specific:

    After open, preserve the returned conversation_id, turn_id, run_id, answer_file_path, log_context, and support_contract for the rest of the …

    line 86
  7. Very specific:

    When result_explanation.show_to_user = true for a blocker or evidence boundary, translate its summary, why, and next_step into one concise …

    line 105
  8. Very specific:

    Do not append it to a successful copy-ready artifact.

    line 106
  9. Very specific:

    When show_to_user = false, do not add extra result-explanation prose.

    line 107
  10. Very specific:

    Do not write result_explanation text into the canonical answer markdown.

    line 108
  11. Distinctive:

    Treat it as repair metadata for the next rewrite/retrace attempt, not as permission to bypass trace or admissibility.

    line 110
  12. Very specific:

    After open, treat the returned status, question_class, question_domain, analysis_origin, route_reason, inner_workflow_id, support_strategy, …

    line 115
  13. Very specific:

    when using public docmason retrieve or docmason trace inside the same canonical ask turn, export each returned log_context field as …

    line 117
  14. Very specific:

    treat compact payloads as the stable host-facing inspection contract.

    line 119
  15. Very specific:

    do not switch to direct Python helpers such as prepare_ask_turn(), complete_ask_turn(), or trace_answer_file(...) as a substitute for the …

    line 132
  16. Very specific:

    After progress returns awaiting-user-decision, present the same gate through the host's native Plan or structured-input UI, without a …

    line 194
  17. Very specific:

    when the current-turn hybrid_refresh_work.json lists render or focus-render assets and the host can inspect images, inspect the relevant …

    line 204
  18. Very specific:

    when a turn is paused in waiting-shared-job, re-enter through hidden open reuse, hidden progress, or hidden finalize.

    line 205
  19. Very specific:

    When the exact final answer is already present at answer_file_path, callers may omit answer_text and use the file-based handshake below.

    line 232
  20. Very specific:

    only completed or boundary permits a final business reply to the user

    line 254
  21. Distinctive:

    If the environment cannot satisfy those capabilities, stop and explain the blocker instead of improvising.

    line 265
  22. Very specific:

    Treat one ordinary user message as one canonical ask turn, and run steps 2 through 4 as the governed open or same-turn reuse phase before …

    line 269
  23. Distinctive:

    keep one live user question mapped to one canonical turn

    line 270
  24. Distinctive:

    when the same live turn and the same active run are re-entered, reuse the existing governed preanswer result instead of restarting …

    line 272
  25. Very specific:

    before calling open, form a thin agent-authored work_brief from the user's actual goal and medium.

    line 276
  26. Very specific:

    when the user explicitly revises a prior result, pass its exact revision_of turn id and the narrow continuation_type, revision_scope, and …

    line 286
  27. Very specific:

    when the user answers a persisted gate, pass the exact resolves_gate_id plus either a legal option_id from that gate or an explicit …

    line 289
  28. Very specific:

    use runtime/bootstrap_state.json as the cached readiness marker

    line 299
  29. Very specific:

    treat workspace-dependent ask as legal only when the prepared environment is self-contained

    line 300
  30. Very specific:

    if the environment is mixed or degraded, let the ask helper repair or surface the governed boundary instead of answering from a partially …

    line 302
  31. Very specific:

    if the environment is ready but no published knowledge base exists yet, route to knowledge-base-sync instead of bluffing a …

    line 304
  32. Distinctive:

    if the published knowledge base is stale but still usable, answer from the published corpus with one concise freshness notice

    line 305
  33. Distinctive:

    if fresh workspace state is genuinely needed, let the ask helper govern prepare or sync rather than improvising it in the workflow

    line 306
  34. Distinctive:

    do not turn simple exact-source asks into answer-critical sync just because pending interaction promotion backlog exists

    line 307
  35. Distinctive:

    when the user has already narrowed to one exact source or unit, treat pending interaction backlog as an advisory notice unless the current …

    line 308
  36. Very specific:

    when workspace freshness depends on live local files, use repo-side live corpus discovery for original_doc/ rather than git-tracked repo …

    line 309
  37. Very specific:

    if prepare or sync becomes a confirmation-required shared job, pause the same turn in awaiting-confirmation

    line 310
  38. Very specific:

    if the same-session turn is already prepared, awaiting-confirmation, or waiting-shared-job, prefer reusing that governed state over …

    line 314
  39. Very specific:

    When governed preanswer returns status = execute, let the canonical ask turn continue through the narrowest ask-routed inner workflow that …

    line 316
  40. Very specific:

    if the same turn is still paused, blocked, or has been routed to workspace-bootstrap or knowledge-base-sync, honor that governed state …

    line 317
  41. Distinctive:

    keep workspace commands sequential inside the live turn

    line 325
  42. Very specific:

    keep the same canonical turn ownership through the inner workflow.

    line 326
  43. Distinctive:

    use published KB artifacts first when they already expose the needed evidence channels

    line 327
  44. Distinctive:

    treat published-KB answering as the ordinary ask path.

    line 328
  45. Very specific:

    treat the published KB as the primary evidence surface: inspect retrieved text, structure, notes, media, and artifact metadata first, then …

    line 329
  46. Very specific:

    start execution with a short support ledger derived from support_contract:

    line 330
  47. Distinctive:

    keep approximate or unresolved reference notices explicit

    line 335
  48. Distinctive:

    treat user-named files and host attachments as current turn evidence by default, but keep their role, authority, freshness, hash and …

    line 338
  49. Distinctive:

    if published artifacts are still insufficient because of hard-artifact semantic gaps, let the canonical routed path enter one governed …

    line 356
  50. Very specific:

    after a covered settlement, rerun retrieve and trace exactly on the post-refresh evidence.

    line 358
  51. Distinctive:

    if that governed path becomes a shared wait or blocked boundary, keep the same turn paused or committed through the existing ask …

    line 359
  52. Very specific:

    write only the final answer under runtime/answers/<conversation_id>/<turn_id>.md

    line 361
  53. Very specific:

    keep scratch work under runtime/agent-work/ when needed

    line 362
  54. Very specific:

    follow the finalize handshake in Canonical Ask Contract: run the final trace on the exact answer-file version, then call hidden finalize, …

    line 363
  55. Very specific:

    if the first finalize attempt returns status = execute with a repairable support_fulfillment, keep the same turn open, do one …

    line 364
  56. Distinctive:

    do not open a second or unbounded repair loop.

    line 365
  57. Distinctive:

    Do not require the user to name a skill or repository command before this workflow can run.

    line 382
  58. Distinctive:

    Do not turn natural-language routing into large keyword tables or a growing odd-question taxonomy.

    line 383
  59. Distinctive:

    Do not create a replacement turn when the same-turn confirmation or waiting path is available.

    line 384
  60. Distinctive:

    Do not relabel approximate references as exact.

    line 385
  61. Very specific:

    Do not treat external-source-verified, model-knowledge, or governed-boundary outcomes as product failures just because the KB path was not …

    line 386
  62. Very specific:

    If published KB artifacts already satisfy the evidence need, do not reopen original_doc/ or rerender source files by reflex.

    line 387

“Common” is measured: kept by at least 6 in 10 of the 60 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Ask

ask is the canonical skill at skills/canonical/ask/SKILL.md. It is the user-facing top-level workflow for ordinary natural-language requests. Use this skill as the default top-level workflow for a new ordinary user request in this workspace unless the request is clearly explicit operator work.

The user should be able to ask naturally. They should not need to name internal workflow IDs or repository mechanics first.

If the request is clearly an explicit setup, status, sync, adapter-maintenance, or operator-review request, switch to the matching top-level workflow instead of forcing it through ask.

Front-Door Law

  • Reading this skill is not legal ask execution.
  • Native-thread reconciliation is not legal ask execution.
  • ask itself remains the only ordinary natural-language front door, and a request counts as ordinary front-door execution only after it has been opened as a canonical ask turn under this workflow.
  • Plan Mode or a similar host planning/collaboration mode should not discourage the host from opening canonical ask for an ordinary user request; if the current mode still blocks ask, the host should use a concise in-band alert or approval question and continue the same user task in the most natural available way rather than silently defaulting to operator-direct or side-path work.
  • See Canonical Ask Contract below for the exact ordinary-ask opening and completion rules. Generated adapter guidance may restate that contract, but it does not replace this workflow surface, and compatible hosts must use that supported entry surface rather than helper stitching, ad hoc probing, or direct lifecycle calls.
  • Direct evidence commands such as public retrieve or trace remain legal operator tools, but they do not complete the ordinary ask contract by themselves.

ask owns front-door legality, same-turn governance, workspace gating, and routing. The routed inner workflow owns the deeper evidence loop.

Turn Terms

  • native turn means the host's own chat turn before DocMason opens canonical ask handling.
  • canonical ask turn means the governed DocMason turn for the current request.
  • runtime ownership means the current request has been opened into that canonical ask turn and is now governed by DocMason.
  • native ledger means host-side audit capture that is not canonical ask truth by itself.
  • interaction-ingest means the runtime holding area for reconciled host activity before any governed promotion.
  • A practical sign that canonical ask really opened is that the request leaves linked runtime artifacts rather than only a host-visible reply, typically under runtime/answers/, runtime/runs/, and runtime/logs/.

Canonical Ask Contract

  • This section is the authoritative ordinary-ask execution contract for compatible hosts.
  • workflow.json remains routing metadata. It does not define the executable host call contract.
  • A request counts as an opened canonical ask turn only when both are true:
    • the supported ask entry surface has returned stable conversation_id, turn_id, run_id, answer_file_path, and log_context
    • the underlying turn has been upgraded to front_door_state = canonical-ask
  • For compatible-host execution, the supported ask entry surface is the hidden host wrapper:
cat <<'JSON' | ./.venv/bin/python -m docmason _ask
{ ...payload... }
JSON
  • Hidden wrapper actions:
    • open
    • progress
    • finalize
  • open request envelope:
{
  "action": "open",
  "question": "<user question>",
  "host_provider": "<provider>",
  "host_thread_ref": "<stable host thread ref>",
  "host_identity_source": "<host identity source>",
  "semantic_analysis": {
    "question_class": "answer",
    "question_domain": "workspace-corpus",
    "route_reason": "<one concise reason>"
  }
}
  • Minimal host hinting rule:
    • semantic_analysis is best-effort. For an ordinary native Codex ask, question_class, question_domain, and one concise route_reason are usually enough.
    • classify by deliverable and evidence basis, not by perceived difficulty:
      • question_class chooses workflow shape. It is not a difficulty score: use answer for a direct answer, explanation, or source-backed summary; use composition for a new evidence-backed work product or synthesis; use retrieval, provenance, or runtime-review only for explicit evidence, citation, or runtime-review requests.
      • question_domain chooses evidence basis. It is not a mirror of question_class: prefer workspace-corpus, external-factual, or general-stable when one of those is the real basis, and reserve composition for composition-shaped evidence planning rather than setting it merely because question_class = composition.
      • a compare, draft, or plan request over workspace materials is therefore usually question_class = composition with question_domain = workspace-corpus.
    • open normalizes missing supported routing fields, derives defaults such as support_strategy, and may refine reference resolution or workspace notices from repository truth.
  • Native Codex fast path:
    • for an ordinary request on the native Codex path, once repo-local .venv is available, call hidden open directly with best-effort semantic_analysis
    • prefer the real native CODEX_THREAD_ID identity from the execution environment; do not hand-fill placeholder host thread references such as codex-desktop-thread or codex-native-thread
    • do not read other workflow skills, workspace-status, workspace-bootstrap, source search, implementation source, or tests first just to decide whether canonical ask may open, whether a named source exists, or which semantic_analysis fields are accepted
    • use those surfaces only after open returns a governed blocker, waiting state, or explicit operator route
  • After open, preserve the returned conversation_id, turn_id, run_id, answer_file_path, log_context, and support_contract for the rest of the same canonical ask turn.
  • Hidden wrapper status meanings:
    • execute means canonical ask is open and the host should continue through the chosen inner workflow.
    • awaiting-confirmation means pause the same turn and wait for the user's confirmation reply.
    • awaiting-user-decision means pause the current work for a material decision that only the user can authorize. It never implies a timeout, default, or skip.
    • waiting-shared-job means pause the same turn and wait for governed shared-job settlement.
    • completed means the turn is committed and a final business answer may be returned.
    • boundary means the turn is committed as a governed boundary and that boundary reply may be returned.
    • blocked means no final business answer may be returned yet.
  • Hidden wrapper next_step is a derived convenience field and should stay aligned with that status law:
    • execute -> continue-inner-workflow
    • awaiting-confirmation -> wait-for-user-confirmation
    • awaiting-user-decision -> wait-for-user-decision
    • waiting-shared-job -> wait-for-shared-job
    • completed -> return-final-answer
    • boundary -> return-boundary-answer
    • blocked -> do-not-return-final-answer
  • Hidden wrapper result_explanation is a derived convenience field, not a new truth surface.
    • When result_explanation.show_to_user = true for a blocker or evidence boundary, translate its summary, why, and next_step into one concise user-facing explanation in the user's language.
    • Do not append it to a successful copy-ready artifact; successful completion returns only the exact business answer plus the separate user_status_line.
    • When show_to_user = false, do not add extra result-explanation prose.
    • Do not write result_explanation text into the canonical answer markdown.
  • Hidden wrapper admissibility_repair is present only for same-turn repairable finalize failures.
    • Treat it as repair metadata for the next rewrite/retrace attempt, not as permission to bypass trace or admissibility.
  • In compatible-host execution, open or same-turn reuse already performs governed preanswer work:
    • question classification and support-strategy selection
    • workspace gating and knowledge-base freshness checks
    • initial inner-workflow routing plus any confirmation or waiting-state settlement
  • After open, treat the returned status, question_class, question_domain, analysis_origin, route_reason, inner_workflow_id, support_strategy, reference_resolution, source_scope_policy, support_contract, and notices as the source of truth for the next step. Do not re-derive them from side-path probing before honoring that result.
  • Retrieve / trace binding rule:
    • when using public docmason retrieve or docmason trace inside the same canonical ask turn, export each returned log_context field as DOCMASON_<FIELD> and then call the public command normally
    • in chat-host execution, prefer --json --compact for interactive inspection; if full nested retrieve or trace detail is genuinely needed, redirect full --json to a local file and inspect it selectively instead of loading the raw payload straight into the live chat context
    • treat compact payloads as the stable host-facing inspection contract; do not rebuild alternate schemas with ad hoc jq assumptions such as .matches
    • compact retrieve inspection should normally start from:
      • session_id
      • results
      • reference_resolution
      • source_scope_policy
    • compact trace inspection should normally start from:
      • trace_id
      • session_id
      • answer_state
      • reference_resolution
      • source_scope_policy
      • issue_codes
    • do not switch to direct Python helpers such as prepare_ask_turn(), complete_ask_turn(), or trace_answer_file(...) as a substitute for the supported path
  • Example retrieve / trace binding:
export DOCMASON_CONVERSATION_ID="<conversation_id>"
export DOCMASON_TURN_ID="<turn_id>"
export DOCMASON_RUN_ID="<run_id>"
export DOCMASON_ENTRY_WORKFLOW_ID="ask"
export DOCMASON_INNER_WORKFLOW_ID="<inner_workflow_id>"
export DOCMASON_FRONT_DOOR_STATE="canonical-ask"

./.venv/bin/python -m docmason retrieve "<query>" --json --compact
./.venv/bin/python -m docmason trace --answer-file "<answer_file_path>" --json --compact
  • progress request envelope:
{
  "action": "progress",
  "conversation_id": "<conversation_id>",
  "turn_id": "<turn_id>",
  "completion_status": "covered | blocked",
  "hybrid_refresh_summary": { "...": "..." }
}
  • A material professional judgment may instead persist a decision gate:
{
  "action": "progress",
  "conversation_id": "<conversation_id>",
  "turn_id": "<turn_id>",
  "decision_frontier": {
    "class": "judgment-authority-gap | high-cost-expression-choice",
    "question": "<material decision question>",
    "why_user_is_required": "<why evidence and reasoning cannot decide it>",
    "evidence_boundary": "<what evidence establishes and cannot decide>",
    "options": [
      {
        "id": "<stable option id>",
        "label": "<short label>",
        "impact": "<real downstream consequence>",
        "recommended": true
      },
      {
        "id": "<stable option id>",
        "label": "<short label>",
        "impact": "<real downstream consequence>",
        "recommended": false
      }
    ],
    "affected_outputs": ["<specific model or artifact scope>"],
    "invalidates": []
  }
}
  • A legal gate has two or three real options, explains their downstream impact, marks exactly one current recommendation, and names the affected output scope. Do not persist a vague or non-material question.
  • After progress returns awaiting-user-decision, present the same gate through the host's native Plan or structured-input UI, without a timeout or default, and wait. If no such UI is available, ask the one material question concisely in-band.
  • The user's answer opens a new linked turn with continuation_type=decision-resolution, the exact gate_id as resolves_gate_id, and either one legal option_id or a non-empty free_form decision. Never settle the old turn in place or reinterpret an unrelated message as consent.
  • completion_status is optional when the caller is only re-entering a waiting-shared-job turn to let the hidden wrapper reconcile deterministic repo-owned shared-job truth.
  • supply completion_status only when the host is actively settling a still-unsettled governed multimodal refresh.
  • when the current-turn hybrid_refresh_work.json lists render or focus-render assets and the host can inspect images, inspect the relevant assets lightly and include render_inspection_used plus inspected_render_assets in hybrid_refresh_summary.
  • when a turn is paused in waiting-shared-job, re-enter through hidden open reuse, hidden progress, or hidden finalize; do not grep runtime/control_plane/ or shared-job files manually.
  • finalize request envelope:
{
  "action": "finalize",
  "conversation_id": "<conversation_id>",
  "turn_id": "<turn_id>",
  "answer_text": "<exact final business answer>",
  "answer_file_path": "<answer_file_path>",
  "response_excerpt": "<short excerpt>",
  "session_ids": ["<selected_session_id>"],
  "trace_ids": ["<selected_trace_id>"],
  "workflow_outcome": {
    "support_basis": "kb-grounded | mixed | external-source-verified | model-knowledge | governed-boundary",
    "session_ids": ["<selected_session_id>"],
    "trace_ids": ["<selected_trace_id>"],
    "bundle_paths": ["<composition bundle path>"]
  }
}
  • answer_text is the preferred terminal input. The wrapper writes it atomically to the canonical answer path, binds the exact digest, reuses only a digest-matching trace, and runs one exact trace when necessary. The successful response keeps answer_text clean and returns one separate user_status_line; internal IDs and support state remain in governance_detail. When the exact final answer is already present at answer_file_path, callers may omit answer_text and use the file-based handshake below.
  • session_ids and trace_ids are optional advanced-caller fields. Omit them when the turn has exactly one ask-owned retrieve session and one final trace candidate. Supply them only when the caller has already selected the canonical pair among multiple ask-owned candidates.
  • workflow_outcome is the preferred finalize-time handoff for workflow-owned facts. Supply it when the inner workflow already knows the correct support_basis, selected session_ids / trace_ids, support-manifest linkage, bundle linkage, or bounded degradation metadata. Older callers may keep using the compatible top-level finalize fields.
  • Legal closure handshake for one canonical ask turn:
./.venv/bin/python -m docmason trace --answer-file "<answer_file_path>" --json --compact

cat <<'JSON' | ./.venv/bin/python -m docmason _ask
{
  "action": "finalize",
  "conversation_id": "<conversation_id>",
  "turn_id": "<turn_id>",
  "answer_file_path": "<answer_file_path>",
  "response_excerpt": "<short excerpt>"
}
JSON
  • Completion rule:
    • only completed or boundary permits a final business reply to the user
    • execute, awaiting-confirmation, awaiting-user-decision, waiting-shared-job, and blocked do not

Required Capabilities

  • local file access
  • shell or command execution
  • ability to inspect structured JSON output
  • ability to inspect rendered images when the answer boundary requires it

If the environment cannot satisfy those capabilities, stop and explain the blocker instead of improvising.

Procedure

  1. Treat one ordinary user message as one canonical ask turn, and run steps 2 through 4 as the governed open or same-turn reuse phase before any inner workflow execution.
    • keep one live user question mapped to one canonical turn
    • reuse the live turn when the same question is continuing
    • when the same live turn and the same active run are re-entered, reuse the existing governed preanswer result instead of restarting preanswer governance
    • return in the user's language unless they ask for another language
  2. Open or reuse the canonical ask turn through the supported path defined in Canonical Ask Contract.
    • reconcile any active native thread, and keep that reconciliation in the native ledger and interaction-ingest path until canonical ask ownership is open
    • before calling open, form a thin agent-authored work_brief from the user's actual goal and medium; include only fields that matter for this turn: deliverable and use, audience, in/out scope, success criteria, distinctions that must survive, named evidence and exemplars, confirmed method/storyline/expression grammar, and high-cost artifact risk
    • simple work may use an empty or very small brief; never ask the user to fill a form and never replace professional framing with keyword inference in deterministic runtime code
    • pass the brief inside best-effort semantic_analysis; keep one concise route_reason, and set needs_latest_workspace_state only when fresh local workspace truth is actually required, and include compact evidence_requirements only when the question needs channel guidance
    • when the user explicitly revises a prior result, pass its exact revision_of turn id and the narrow continuation_type, revision_scope, and affected_outputs; adjacency in the same chat is not sufficient authority to inherit evidence or accepted scopes
    • when the user answers a persisted gate, pass the exact resolves_gate_id plus either a legal option_id from that gate or an explicit non-empty free_form decision; do not reinterpret arbitrary later text as gate approval
    • let open normalize supported routing fields, resolve user-native source references when the user names a document, path, page, slide, sheet, heading, or similar locator, and return the governed turn binding
  3. During open or same-turn reuse, let the governed ask path choose the smallest evidence basis that can support the answer correctly and truthfully before deeper workflow execution.
    • workspace-corpus -> KB-first
    • composition -> KB-first with explicit evidence planning
    • external-factual -> web-first
    • general-stable -> model knowledge when the boundary is explicit
  4. During that same governed preanswer step inside open, check workspace state only when the answer really depends on workspace truth.
    • use runtime/bootstrap_state.json as the cached readiness marker
    • treat workspace-dependent ask as legal only when the prepared environment is self-contained
      • self-contained means the repo-local steady-state runtime is trusted for ordinary ask work
    • if the environment is mixed or degraded, let the ask helper repair or surface the governed boundary instead of answering from a partially trusted runtime
    • allow safe silent bootstrap or repair when the workspace-dependent path can continue safely
    • if the environment is ready but no published knowledge base exists yet, route to knowledge-base-sync instead of bluffing a workspace-grounded answer
    • if the published knowledge base is stale but still usable, answer from the published corpus with one concise freshness notice
    • if fresh workspace state is genuinely needed, let the ask helper govern prepare or sync rather than improvising it in the workflow
    • do not turn simple exact-source asks into answer-critical sync just because pending interaction promotion backlog exists
    • when the user has already narrowed to one exact source or unit, treat pending interaction backlog as an advisory notice unless the current turn truly depends on interaction-derived evidence
    • when workspace freshness depends on live local files, use repo-side live corpus discovery for original_doc/ rather than git-tracked repo search
    • if prepare or sync becomes a confirmation-required shared job, pause the same turn in awaiting-confirmation
      • accept short same-session yes or no replies as approve or decline
        • yes continues the same task
        • no commits the same turn as abstained + governed-boundary
    • if the same-session turn is already prepared, awaiting-confirmation, or waiting-shared-job, prefer reusing that governed state over rerunning shared-state mutation
    • while a turn is in waiting-shared-job or awaiting-confirmation, do not bypass the governed path by answering from original_doc/, knowledge_base/staging/, or .staging-build
  5. When governed preanswer returns status = execute, let the canonical ask turn continue through the narrowest ask-routed inner workflow that matches the ask.
    • if the same turn is still paused, blocked, or has been routed to workspace-bootstrap or knowledge-base-sync, honor that governed state first instead of forcing one of the ask-routed evidence workflows
    • for answer-eligible ask-time execution, the workflow must choose the most appropriate one of the following 5 inner workflows and continue through that path
    • direct supported answer -> grounded-answer
    • evidence-backed drafting, planning, or research -> grounded-composition
    • evidence-only request -> retrieval-workflow
    • provenance or citation request -> provenance-trace
    • runtime review request -> runtime-log-review
  6. Route into the chosen ask-routed inner workflow and let it own the evidence loop.
    • keep workspace commands sequential inside the live turn
    • keep the same canonical turn ownership through the inner workflow; retrieve, trace, raw-source inspection, or internal helper probing may support that live turn, but they do not replace canonical ask ownership or its completion rules
    • use published KB artifacts first when they already expose the needed evidence channels
    • treat published-KB answering as the ordinary ask path; governed Lane B follow-up remains sync/publication-owned and may surface as freshness or degradation notice, but it is not a separate host-driven ask step
    • treat the published KB as the primary evidence surface: inspect retrieved text, structure, notes, media, and artifact metadata first, then inspect cited focus_render_assets or render spans when the question is genuinely visual or layout-sensitive, and only then consider governed refresh or source fallback
    • start execution with a short support ledger derived from support_contract:
      • which source boundary must survive
      • which comparison sources must both survive
      • which published evidence channels are required
      • whether a single contract-repair chance exists for this turn
    • keep approximate or unresolved reference notices explicit
    • normalize the additive thin work_brief, revision lineage, accepted scopes, and turn evidence packet before deciding whether evidence work is necessary
    • treat user-named files and host attachments as current turn evidence by default, but keep their role, authority, freshness, hash and supersession semantics explicit; do not promote them into the durable KB merely because they were used once
    • distinguish evidence-gap, reasoning-uncertainty, and judgment-authority-gap: solve the first two autonomously; persist a decision gate only for the third or for a genuinely expensive expression choice
    • for constraint-update and evidence-neutral decision-resolution, inherit legal evidence, skip sync and retrieval, preserve unrelated accepted scopes, retrace only the exact revised answer or artifact, and finalize
    • record an accepted scope only when the current user explicitly accepted it and the host sets accepted_by_user: true; reopen it only after explicit user authority with reopened_by_user: true. Dependency drift marks a scope at-risk but never silently reopens or replaces it
    • for evidence-refresh or mixed, retrieve only the evidence delta and invalidate only explicitly affected or dependency-linked outputs
    • unresolved relevance alone does not authorize global sync; use task-scoped freshness and expose a boundary when critical target freshness cannot be established
    • let the routed inner workflow own retrieval, trace, render inspection, and answer or composition drafting
    • if published artifacts are still insufficient because of hard-artifact semantic gaps, let the canonical routed path enter one governed narrowed hybrid refresh instead of improvising raw source fallback
      • this ask-owned narrowed hybrid refresh is the Lane C path and is the only ordinary ask follow-up settled through hidden progress
      • after a covered settlement, rerun retrieve and trace exactly on the post-refresh evidence; if the result is commit-admissible but still only partially supported, finalize honestly as partially-grounded instead of starting a second refresh
    • if that governed path becomes a shared wait or blocked boundary, keep the same turn paused or committed through the existing ask control-plane states rather than opening a side path
  7. Complete the turn through the supported completion path.
    • write only the final answer under runtime/answers/<conversation_id>/<turn_id>.md
    • keep scratch work under runtime/agent-work/ when needed
    • follow the finalize handshake in Canonical Ask Contract: run the final trace on the exact answer-file version, then call hidden finalize, passing selected artifact IDs only when the turn is ambiguous
    • if the first finalize attempt returns status = execute with a repairable support_fulfillment, keep the same turn open, do one contract-aware rewrite and retrace, then finalize once more
    • do not open a second or unbounded repair loop; the second finalize attempt must close honestly
    • let the commit barrier run only after the admissibility gate passes
    • preserve answer_state, support_basis, optional support_manifest_path, and linked session or trace IDs
  8. Return the result cleanly.
    • direct answer when supported
    • explicit non-answer boundary when not
    • one concise freshness or waiting note only when it materially helps the user
    • return the business answer_text unchanged, followed outside the artifact by exactly one concise completion line; use user_status_code as the stable meaning and render the line naturally in the user's language (completed-and-verified or completed-with-evidence-boundary)
    • show governance_detail only for blockers, evidence boundaries, diagnostics, or explicit user inspection; never append trace IDs, manifest paths, repair codes, or release-entry notices to a successful copy-ready artifact

Escalation Rules

  • Do not require the user to name a skill or repository command before this workflow can run.
  • Do not turn natural-language routing into large keyword tables or a growing odd-question taxonomy.
  • Do not create a replacement turn when the same-turn confirmation or waiting path is available.
  • Do not relabel approximate references as exact.
  • Do not treat external-source-verified, model-knowledge, or governed-boundary outcomes as product failures just because the KB path was not fully grounded.
  • If published KB artifacts already satisfy the evidence need, do not reopen original_doc/ or rerender source files by reflex.

Completion Signal

  • The workflow is complete when the user question has been routed through one canonical turn, the resulting logs are linked correctly, and the final answer or explicit boundary has been returned cleanly.

Notes

  • A reconciled native turn is still only host-side context until the matching canonical ask turn has been opened.
  • grounded-answer and grounded-composition remain inner specialist workflows.
  • Host Hooks are optional accelerators. Never assume their context injection or one-shot Stop continuation ran; without them, follow the same canonical contract directly.
  • Tracked repo search, live corpus discovery, knowledge-base artifact discovery, and runtime artifact discovery are different surfaces; do not substitute one for another silently.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.