Binder epistemic report

Type: agentic-system-analyses/types/agentic-system-epistemic-report.md

Source-and-claim boundary

The question is how Binder's typed shared memory acquires truth-apt content, checks it and lets it affect later work. Scope and exclusions are see boundary; evidence identities are see Source register. SRC-1 supplies implementation and doctrine/design only. No target execution, provider calls or tests were run. No observed model output, accepted journal instance, supplied run trace or causal comparison establishes compliance, factual accuracy or benefit. External provider weights and host scheduling remain uninspected; this prevents model-internal learning and end-to-end activation conclusions, not source-level route findings.

Public storage/memory claim: EPI-CLM-agent-memory. The source also requests evidence-grounded personal insights: EPI-CLM-period-learnings. These are not interchangeable claims of factual validation. Structural validation and history consistency are see RT-CLM-transaction-boundary; post-commit limits are see RT-CLM-derived-effects.

Entry point or operation	Source path in SRC-1	Coverage and limit
CLI registration and dynamic dispatch	packages/cli/src/index.ts	Direct full read; core commands plus docs/journal/undo and workspace commands; no hidden universal model evaluator inferred.
Create/update/delete; MCP and HTTP mutations; library	packages/cli/src/commands/create.ts, packages/cli/src/commands/delete.ts, packages/cli/src/mcp/tools/index.ts, packages/cli/src/http/server.ts, packages/cli/src/workspace.ts	Direct searches of create/delete and full MCP registry read; remaining implementation supplied by RT-RTE-graph-update. Registry exposes schema/search/transact.
Read/search/schema and document/editor paths	packages/cli/src/commands/read.ts, packages/cli/src/commands/search.ts, packages/cli/src/plugins/docs/commands.ts, packages/cli/src/lsp/handlers/save-handler.ts	Supplied RT-OBJ-knowledge-graph, RT-RTE-document-sync. Extraction fidelity not independently audited; prevents a lossless semantic round-trip conclusion.
Transaction import, preview, rollback, squash, export/log	packages/cli/src/commands/transaction.ts	Direct handler excerpt and operation search; EPI-RTE-import-preview. Squash internals and complete export encoding uninspected: no system-complete lineage-preservation verdict.
Undo/redo and journal repair	packages/cli/src/plugins/undo/commands.ts, packages/cli/src/plugins/journal/commands.ts	Full undo/redo read and journal command search; EPI-RTE-undo-redo, supplied RT-RTE-journal-repair.
Hooks, workspace scripts/plugins, custom HTTP modules	packages/repo/src/local/plugin-loader.ts, packages/cli/src/http/server.ts	Supplied RT-RTE-extension-load. Arbitrary user code not captured; prevents universal admission guarantees.
Git importer and generic LLM helper	examples/journal/git-import-logs.ts, packages/cli/src/lib/llm.ts	Direct full importer read; EPI-RTE-import-format, supplied RT-RTE-git-import, RT-RTE-llm-call. No core helper consumer established.
Import/modeling/app skills	skills/binder-import/SKILL.md, skills/binder-modeling/SKILL.md, skills/binder-app/SKILL.md	Full import read; scoped search of modeling/app checks. EPI-RTE-import-mapping. Remaining modeling/app instructions not fully assessed; prevents complete prompt-policy coverage.
Journal summary, learning and planning	examples/journal/skills/day-summary/SKILL.md, examples/journal/skills/period-summary/SKILL.md, examples/journal/skills/period-update/SKILL.md, examples/journal/skills/day-next-plan/SKILL.md	Direct full reads; EPI-RTE-journal-review. Other journal skills scoped searches only; no comprehensive journal-policy negative.
Task capture/refinement/investigation/implementation	examples/workflow/skills/capture-task/SKILL.md, examples/workflow/skills/refine-backlog/SKILL.md, examples/workflow/skills/implement-task/guides/investigate.md	Direct full reads; EPI-RTE-task-findings, supplied RT-RTE-task-guide. Design/discuss/write-tests guides not independently read; no full theory-cycle verdict.
Authored evaluators	packages/cli/tests/doc-sync.test.ts, packages/repo/src/knowledge-graph.test.ts	Boundary's supplied test-source inspection only, not a passing run or factual-content evaluator.

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence	Gap/limit
RT-OBJ-knowledge-graph	Caller fields may assert facts; schema/IDs alone need not	see record	SRC-1, packages/repo/src/knowledge-graph.ts	Container is not one proposition or one warrant class.
EPI-OBJ-log	Dated activity descriptions	see record	SRC-1, examples/journal/git-import-logs.ts	No observed generated instance.
EPI-OBJ-learnings	Personal insights, summaries, achievements and imported information	see record	SRC-1, examples/journal/skills/period-summary/SKILL.md	Separate arithmetic, historical reports and inferred patterns within the record.
EPI-OBJ-task	Bug explanations and reproduction claims; requirements also include non-truth-apt goals	see record	SRC-1, examples/workflow/skills/implement-task/guides/investigate.md	No retained investigation instance or test result.
RT-CMP-journal-model, RT-CMP-llm-model	Not determinable for internal distributed-parametric content	see records	SRC-1, examples/journal/git-import-logs.ts, packages/cli/src/lib/llm.ts	Provider parameters, training, endpoint resolution and internal processing uninspected; observable wrapper policy remains assessable.

Authority-route ledger

All owned fields below say see their declaring record; the short local ledger judgments do not replace those fields. Distinct rows separate checking from disposition, and retention from integration. No acceptance of an actual ampliative proposition is established.

Route ID	Route function	Architectural status	Object/candidate ID	Content/update relation	Transition or check target	Evaluator/condition and domain	Activation and timing	Possible or observed result	Implemented force	Epistemic authority and scope	Operational authority	Behavioral-authority path	Evidence	Claim IDs	Mismatch marker	Gap/limit
RT-RTE-graph-update	check/evidence production	implemented	RT-OBJ-knowledge-graph	no content change	see route	see route	see route	see route	see route	Structural/schema and sequence license only	see route	see route	SRC-1, packages/repo/src/transaction-processor.ts	RT-CLM-transaction-boundary	none	Factual truth is outside this evaluator's target.
RT-RTE-graph-update	retention	implemented	RT-OBJ-knowledge-graph	truth-apt transformation: acquisition/import	see route	see route	see route	see route	see route	External source warrant unknown; persistence is not endorsement	see route	see route	SRC-1, packages/repo/src/knowledge-graph.ts	EPI-CLM-agent-memory	none	Content may instead be policy/config; classification applies to truth-apt fields only.
EPI-RTE-import-preview	check/evidence production	implemented	RT-OBJ-knowledge-graph	no content change	see route	see route	see route	see route	see route	Preview of processability, not truth or guaranteed batch success	see route	see route	SRC-1, packages/cli/src/commands/transaction.ts	none	none	Warning continues and final return can be success.
RT-RTE-journal-repair	lineage/freshness/recovery	implemented	RT-OBJ-knowledge-graph	no content change	see route	see route	see route	see route	see route	Log continuity only	see route	see route	SRC-1, packages/cli/src/plugins/journal/sync.ts	RT-CLM-transaction-boundary	none	May replace graph state without fresh factual review.
EPI-RTE-undo-redo	lineage/freshness/recovery	implemented	RT-OBJ-knowledge-graph	no content change	see route	see route	see route	see route	see route	Restoration does not restore factual warrant	see route	see route	SRC-1, packages/cli/src/plugins/undo/commands.ts	none	none	Withdrawal is caller-driven, not automatic claim refutation.
RT-RTE-git-import	content transformation	implemented	EPI-OBJ-log	truth-apt transformation: indeterminate	see route	see route	see route	see route	see route	Candidate only	see route	see route	SRC-1, examples/journal/git-import-logs.ts	none	none	Requested compression may preserve or add meaning; no output distinguishes them.
EPI-RTE-import-format	content transformation	implemented	EPI-OBJ-log	truth-apt transformation: non-ampliative reshaping	see route	see route	see route	see route	see route	No additional source warrant	see route	see route	SRC-1, examples/journal/git-import-logs.ts	none	none	Concrete trace in record; not a target execution.
RT-RTE-llm-call	check/evidence production	implemented	RT-CMP-llm-model	no content change	see route	see route	see route	see route	see route	Requested response shape only	see route	see route	SRC-1, packages/cli/src/lib/llm.ts	none	none	Schema validity does not establish content truth; no downstream consumer found.
EPI-RTE-import-mapping	disposition/acceptance	doctrine only	RT-OBJ-knowledge-graph	truth-apt transformation: acquisition/import	see route	see route	see route	see route	see route	Mapping and operational import approval, source warrant unknown	see route	see route	SRC-1, skills/binder-import/SKILL.md	none	none	No source-truth criterion or observed approval.
EPI-RTE-journal-review	check/evidence production	doctrine only	EPI-OBJ-learnings	truth-apt transformation: indeterminate	see route	see route	see route	see route	see route	Personal evidence and correction, not general mechanism validation	see route	see route	SRC-1, examples/journal/skills/period-summary/SKILL.md	EPI-CLM-period-learnings	none	Inferred patterns can exceed logs; specific non-entailment not established without a candidate.
EPI-RTE-journal-review	disposition/acceptance	doctrine only	EPI-OBJ-learnings	no content change	see route	see route	see route	see route	see route	User-confirmed account for journal/planning use only	see route	see route	SRC-1, examples/journal/skills/day-summary/SKILL.md	EPI-CLM-period-learnings	none	Actual evidence-consuming decision unobserved.
EPI-RTE-journal-review	retention	doctrine only	EPI-OBJ-learnings	no content change	see route	see route	see route	see route	see route	Confirmation is not external validation	see route	see route	SRC-1, examples/journal/skills/period-summary/SKILL.md	EPI-CLM-agent-memory	none	Graph write affordance is implemented separately.
EPI-RTE-journal-review	lifecycle integration	doctrine only	EPI-OBJ-learnings	no content change	see route	see route	see route	see route	see route	Intended confirmed summaries elevated through periods	see route	see route	SRC-1, examples/journal/skills/period-summary/SKILL.md	EPI-CLM-period-learnings	none	Conditional on prior acceptance; arbitrary graph records have no such prerequisite.
EPI-RTE-journal-review	behavior/policy adaptation	doctrine only	EPI-OBJ-learnings	non-truth-apt policy/content update: agreed plans and goals	see route	see route	see route	see route	see route	Plans are choices, not accepted factual theories	see route	see route	SRC-1, examples/journal/skills/period-update/SKILL.md	none	none	Direct adaptation route can operate without produced knowledge.
EPI-RTE-task-findings	content transformation	doctrine only	EPI-OBJ-task	truth-apt transformation: indeterminate	see route	see route	see route	see route	see route	Candidate explanation and reproduction account	see route	see route	SRC-1, examples/workflow/skills/implement-task/guides/investigate.md	none	none	Reproduction verifies symptom, not uniquely the proposed cause.
RT-RTE-task-guide	operational admission/selection/consumption	doctrine only	EPI-OBJ-task	no content change	see route	see route	see route	see route	see route	Task design contract and conditional test reference	see route	see route	SRC-1, examples/workflow/skills/implement-task/SKILL.md	none	none	Guide selection can rely on field presence before any epistemic acceptance.

System-claim versus route comparison

Claim ID	Claimed operation or warrant	Claim source and layer	Doctrine/design support	Implemented routes	Observed-run support	Causal support and limits	Supported conclusion	Mismatch/unknown
EPI-CLM-agent-memory	see record	SRC-1, README.md; doctrine/design	Shared typed state	RT-RTE-graph-update	None supplied	None supplied	Typed persistence/read-back is wired	Transparency and easy maintenance not measured; no claim of universal truth checking implied.
EPI-CLM-period-learnings	see record	SRC-1, examples/journal/skills/period-summary/SKILL.md; doctrine/design	Grounded proposals, user corrections, retained learnings	Underlying RT-RTE-graph-update; interpreting policy EPI-RTE-journal-review is doctrine only	None supplied	No intervention or comparison	A human-assisted evidence-responsive procedure is specified	No demonstrated accepted ampliative claim, transfer or future capacity gain.
RT-CLM-transaction-boundary	see record	SRC-1, packages/repo/src/transaction-applier.ts; implementation	see record	RT-RTE-graph-update, EPI-RTE-undo-redo	see record	No factual or host-wide license	Structural admission and continuity have bounded force	Prepared apply bypasses fresh preparation; plugins can cross effect boundary.

Bounded conclusion

Binder implements storage, retrieval and structural admission of caller-supplied state, not a universal factual acceptance mechanism. Acquisition preserves accessible text/fields as graph data, but source warrant is unknown unless a caller supplies and checks it; attribution metadata alone does not authenticate authorship. Prepared replay, undo/redo and caller deletion can withdraw or restore operational state without judging the facts it states. These limits are see RT-RTE-graph-update, RT-RTE-journal-repair, EPI-RTE-undo-redo.

The journal importer requests model-generated activity summaries and deterministically reshapes the returned list into insertion mutations. It only prints a proposal. Its actual prompt omits commit hashes and time-of-day; therefore its request for actual timestamps lacks that source input. Shape parsing and timestamp positioning cannot recover the omitted evidence. No measured deduplication or factual correctness follows; see EPI-RTE-import-format. Arithmetic averages in journal prompts describe entailed calculations conditional on trustworthy scores and correct execution, not an inspected numerical evaluator or a warranted conclusion about productivity.

Human-mediated journal procedures do specify evidence, criteria and intended use: compare records and user corrections, confirm the personal account, then save it for journal review and subsequent planning. Import approval instead licenses mapping/application, and task readiness licenses work selection; neither is automatically epistemic acceptance. Proposed learnings may import outside information or infer patterns; without concrete outputs their relation remains indeterminate. The architecture requests checking, correction, retention and conditional integration, but no observed acceptance of an ampliative claim is established. See EPI-RTE-journal-review, EPI-RTE-import-mapping, EPI-RTE-task-findings.

Operational authority is heterogeneous. Graph validation enforces mutation admission immediately; restored state changes subsequent queries. Journal and task text advise external agents through skill bodies and requested reads over later workflow invocations; actual compliance is unobserved. Plans can change directly from activity/user discussion without a truth-apt intermediate. Test success is a bounded outcome check, not proof of root-cause attribution or transfer. The uninspected guides and extension behavior prevent a system-complete negative. No single system epistemic status follows, and persistence, reflection-themed prose, learning-field names, model opacity and autonomous host execution establish none of learning benefit, a completed theory criticism cycle or exercised self-improvement by themselves.

Shared records

Operative objects

EPI-OBJ-log

Label: Journal importer log candidates

Source-native identity: logEntries and changeset.log. Natural-language dated activity strings inside symbolic insertion mutations; process memory then stdout JSON, not newly persisted by the importer. Distinct from the retained graph container; output may replace/add its content but is not presently a part of that container. Closest supplied record: RT-RTE-git-import, which supplies the whole production route, not an object declaration. SRC-1, examples/journal/git-import-logs.ts.

const logEntries = parseLogEntries(llmOutput); --- examples/journal/git-import-logs.ts

EPI-OBJ-learnings

Label: Journal summaries and learnings

Part of: RT-OBJ-knowledge-graph

Source-native identity: journal summary, learnings, achievements and score fields. Text is natural-language interpreted by people/models; numeric scores are symbolic inputs to requested averaging. Storage substrate: graph fields in SQLite when written; model proposals first inhabit host dialogue. These are distinct parts of the runtime container with human correction and later journal consumers, not an additional store. Scores encode judgments rather than independent measurements. SRC-1, examples/journal/skills/day-summary/SKILL.md, examples/journal/skills/period-summary/SKILL.md.

Learnings must be grounded in period evidence: personal insights and information from external sources like articles, not generic productivity advice --- examples/journal/skills/period-summary/SKILL.md

EPI-OBJ-task

Label: Task diagnosis and design fields

Part of: RT-OBJ-knowledge-graph

Source-native identity: rootCause, stepsToReproduce, specification, acceptanceCriteria, task status and links. Textual diagnosis/design fields receive consequences through model interpretation; status/link identifiers also select symbolic query results. Storage substrate: graph fields after updates, with source files in the host filesystem. Distinct from journal parts by target, check and consumer. Closest supplied RT-RTE-task-guide owns whole task progression. SRC-1, examples/workflow/skills/implement-task/guides/investigate.md, examples/workflow/skills/capture-task/SKILL.md.

For rootCause, explain why the bug happens, not just where. For sourceFiles, list all files involved — not just the entry point. --- examples/workflow/skills/implement-task/guides/investigate.md

Routes

EPI-RTE-import-preview

Label: Transaction import dry-run and admission

Distinct from RT-RTE-graph-update: this CLI caller has an explicit nonmutating preview branch and a separate sequential update branch. SRC-1, packages/cli/src/commands/transaction.ts. Caller supplies files/stdin and selection; command parses, previews via kg.process or applies via kg.update. Preview targets current structural processability, not factual truth; warnings are printed and processing continues. Apply may fail on a later transaction after earlier ones committed. No automatic per-content human approval is imposed. Operating mode is open import, no answer oracle; CLI metadata and source tags persist on actual admission. Recovery is graph rollback/undo, not whole-file atomicity.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: printed previews/warnings and success in dry-run; transactions or failure during apply.
• Later read-back: applied records are subsequently queryable; previews themselves do not change graph state.
• Delegated visibility: requesting caller sees CLI output; other workspace clients can read applied records.
• Selection predicate: chosen files/stdin, selection options and dryRun flag.
• Invalidation or expiry: subsequent graph state can invalidate a preview; rollback can remove applied suffixes.
• Activation or effect: process results advise the caller; update success enforces structural admission and persistence.
• Evidence limits: implementation only; preview success is neither factual acceptance nor guaranteed batch application.

ui.warning(Transaction ${i + 1} would fail:); --- packages/cli/src/commands/transaction.ts

ui.block(() => ui.info("Dry run complete - no changes made")); --- packages/cli/src/commands/transaction.ts

EPI-RTE-undo-redo

Label: Undo withdrawal and prepared redo

Distinct from supplied RT-RTE-journal-repair: caller-selected step count, undo-log source and redo's current-hash check, rather than log-authoritative divergence repair. SRC-1, packages/cli/src/plugins/undo/commands.ts. Handler validates count, fetches display transactions, rolls back selected suffix; redo reads parsed undo-log transactions, checks predecessor and calls kg.apply sequentially. Software can veto count/history mismatch; user chooses restoration, not per-proposition factual endorsement. Open recovery mode, no answer oracle. Guidance is command contract and stored history; graph state/history changes, not arbitrary external side effects.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: displayed transactions and success or count/history/application error.
• Later read-back: subsequent queries see rolled-back/restored graph state; staged transactions can feed redo.
• Delegated visibility: caller sees display; workspace readers and graph subscribers see state effects.
• Selection predicate: steps, current version, undo-log tail and predecessor hash.
• Invalidation or expiry: changed current hash blocks redo of the captured predecessor; further updates supersede state.
• Activation or effect: rollback withdraws graph mutations; prepared apply restores transactions without fresh-input processing.
• Evidence limits: implementation only; no automatic semantic refutation, factual reapproval or multi-step atomicity established.

"Cannot redo: repository state has changed since undo", --- packages/cli/src/plugins/undo/commands.ts

const applyResult = await kg.apply(tx); --- packages/cli/src/plugins/undo/commands.ts

EPI-RTE-import-format

Label: Git prompt omission and insertion formatting

Part of: RT-RTE-git-import

The importer code owns Git selection, prompt formatting, line extraction and mutation positions; external model owns candidate wording. Context is Git output plus prior/current log strings. SRC-1, examples/journal/git-import-logs.ts. This part distinguishes fixed transformations from uninspected semantic generation. The target of parseTime is a string prefix, not historical timing; parseLogEntries checks a list-line prefix, not YAML validity or truth. It has no semantic veto or answer oracle. The model prompt asks for meaningful, nonoverlapping entries, but these criteria are not checked by the parser. Open date-based import; no internal graph write or content approval.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: prompt text, extracted strings, insertion mutations and ultimately printed changeset, or empty-extraction failure.
• Later read-back: inapplicable — this script does not persist its proposal for another invocation.
• Delegated visibility: filtered prompt reaches external CLI; caller receives parsed proposal rather than raw commit evidence.
• Selection predicate: date, message exclusion regexes, lines beginning with a dash and space, parsed time ordering.
• Invalidation or expiry: temporary prompt deleted after model completion; no proposal expiry or enforced semantic deduplication.
• Activation or effect: code can order insertions; absent entries cause failure, not acceptance or storage.
• Evidence limits: static deduction only; no model output or source accuracy observation. Shape processing adds no warrant.

Concrete deterministic trace: a Git log line abc|2026-10-14|Fix parser becomes prompt line - 2026-10-14 Fix parser if not excluded and no stat lines survive. Hash abc is discarded, and --date=short supplies no HH:MM. Stat extraction retains only lines whose trimmed text starts with |, so ordinary path | 4 ++-- lines do not survive. A returned line - "09:00 - Fixed parser" yields string 09:00 - Fixed parser; against existing 10:00 - Review, minute values 540 and 600 yield ["insert", "09:00 - Fixed parser", 0]. That is non-ampliative reformatting/position selection of the candidate text, not evidence that work occurred at 09:00. [] extracts zero entries and fails despite being allowed in the prompt. A counterexample to a blanket no-check claim is the subprocess exit and empty-list check; neither checks historical accuracy.

const GIT_DATE_FORMAT = "short"; --- examples/journal/git-import-logs.ts

return - ${date} ${msg}${changes.length > 0 ?\n Files: ${changes.join(", ")}: ""}; --- examples/journal/git-import-logs.ts

.filter((line: string) => line.trim().startsWith("| ")) --- examples/journal/git-import-logs.ts

const entry = trimmed.slice(1).trim().replace(/^"/, "").replace(/"$/, ""); --- examples/journal/git-import-logs.ts

EPI-RTE-import-mapping

Label: Human-approved import mapping

No counterpart after comparison with RT-RTE-graph-update: this is a shipped agent procedure before executable graph admission, not another write implementation. SRC-1, skills/binder-import/SKILL.md. User triggers import; host agent samples source, inspects schema, proposes mapping and duplicates, generates batches, previews and asks approval to apply. Human can veto mapping/batches; graph software independently rejects malformed writes. Intended use is mapped operational records; criterion is compatible meaning/types, disclosed skipped fields and chosen duplicate handling, not independent truth verification. Stable-key upserts substitute for per-record search; fuzzy matches require user decisions. Operating mode open import, no answer oracle. Recovery requested via undo; generated YAML and tracking support resumption.

• implementation conclusion status: claimed
• operation conclusion status: uninspected
• Immediate return: requested mapping, deduplication report and batch results.
• Later read-back: imported graph records become search/read inputs; YAML retained until import verified.
• Delegated visibility: user sees proposal/preview; host gets source samples and schema, not necessarily entire large source.
• Selection predicate: source format/size, natural key, batch size, dependency order and user approval.
• Invalidation or expiry: schema gaps pause import; subsequent updates/undo change admitted records.
• Activation or effect: natural-language policy requests human approval; underlying CLI can operate without this procedure.
• Evidence limits: doctrine only; source warrant unknown, no executed approval or lossless conversion established.

Wait for user confirmation before proceeding. --- skills/binder-import/SKILL.md

Only after user approval: --- skills/binder-import/SKILL.md

EPI-RTE-journal-review

Label: Journal evidence review and later planning

Distinct from RT-RTE-git-import, which prints candidate log insertions, and RT-RTE-task-guide, which targets product work. SRC-1, examples/journal/skills/day-summary/SKILL.md, examples/journal/skills/period-summary/SKILL.md, examples/journal/skills/period-update/SKILL.md, examples/journal/skills/day-next-plan/SKILL.md. External host agent reads target/related entries, proposes summaries/scores/insights; person corrects/confirms; agent is instructed to save accepted journal text and carry agreed work forward. Criterion is consistency with activity records and user corrections for personal journal/planning scope. Child completion signals and explicit abandonment classify done/next/drop; unlogged work is not automatically nonoccurrence. User can veto/correct; software admission checks structure separately. Operating mode open personal review, no independent answer oracle. Prompt guidance persists as static skill; confirmed fields and plans can persist via Binder; no durable confirmation/evidence-link record is required by these prompts.

• implementation conclusion status: claimed
• operation conclusion status: uninspected
• Immediate return: proposed reflection, scores, uncertain items and confirmed update instructions.
• Later read-back: period-summary reads children's learnings/summaries; period-update reads retained learning lines; day-next-plan reads source summary and achievements to decide future plan.
• Delegated visibility: selected graph fields and conversation reach external host agent/user; automatic extra context depends on uncaptured extension.
• Selection predicate: period/day offset, parent/children relations, recent entries, incomplete items and human corrections.
• Invalidation or expiry: confirmation can revise/remove proposals; later updates replace fields; explicit drop prevents carry-forward.
• Activation or effect: advisory interpretation and prompt-requested writes/plan changes, not runtime-enforced approval or measured behavior change.
• Evidence limits: doctrine only; no accepted claim instance, observed write, reliable arithmetic execution or causal capacity improvement.

Gaps may mean "not logged yet" rather than "didn't happen" — ask "anything not captured yet?" rather than assuming omission. --- examples/journal/skills/day-summary/SKILL.md

Correct based on my input. Ask at most one follow-up question, only for ambiguous status decisions. If I say "looks good," apply defaults and write. --- examples/journal/skills/period-summary/SKILL.md

Confirm, then write. --- examples/journal/skills/period-update/SKILL.md

Flag zombie tasks — anything that's been unfinished for 3+ consecutive days: --- examples/journal/skills/day-next-plan/SKILL.md

EPI-RTE-task-findings

Label: Task capture refinement and diagnosis

This spans preimplementation capture/refinement and the investigation part of RT-RTE-task-guide; not identical to that supplied whole guide-selection route and not wholly contained by it. SRC-1, examples/workflow/skills/capture-task/SKILL.md, examples/workflow/skills/refine-backlog/SKILL.md, examples/workflow/skills/implement-task/guides/investigate.md. User input is acquired; agent labels inferred fields, screens duplicates, proposes readiness and links, then requests approval. Readiness is sufficient description/constraints/testable criteria, not accepted explanatory truth. Investigation instead asks the agent to reproduce a bug, trace code, explain why and write diagnosis fields, then remove debug edits. Reproduction target is the reported symptom under recorded steps; root-cause target is the explanation of that symptom. The former is requested as an environment check; the latter lacks an independently enforced discriminating intervention. Human veto is explicit for capture/refinement, not per diagnosis update. Guidance is source, task and skill; open task work, no universal answer oracle. Common graph rollback applies to fields; cleanup asks to remove temporary code edits, not the diagnosis.

• implementation conclusion status: claimed
• operation conclusion status: uninspected
• Immediate return: proposed/updated task or diagnosis, with debugging changes requested to be removed.
• Later read-back: task reread and nonempty diagnosis/design fields select subsequent guide/implementation through the supplied runtime route.
• Delegated visibility: task content, project source and guide text reach host agent; user sees approval proposals.
• Selection predicate: user task, missing fields, duplicate/overlap assessment, DoR and reproduction results.
• Invalidation or expiry: user corrections, cancellation/merge and later task updates can replace fields; no automatic explanation invalidation found in inspected files.
• Activation or effect: prompt-directed retention and guide selection; actual execution, approval and code effects unobserved.
• Evidence limits: doctrine only; symptom reproduction does not identify a unique mechanism or establish transfer.

Propose changes to the user and wait for approval before executing --- examples/workflow/skills/refine-backlog/SKILL.md

1. Reproduce. Use stepsToReproduce if present. Otherwise, derive reproduction steps from description and details. Run the code and confirm the bug exists. --- examples/workflow/skills/implement-task/guides/investigate.md

Remove any temporary changes you made during investigation (debug prints, scratch files). The goal is findings, not code changes. --- examples/workflow/skills/implement-task/guides/investigate.md

Claims

EPI-CLM-agent-memory

Label: Typed persistent agent memory

Claimed operation: typed transparent persistent shared agent/person state, not universal factual endorsement. SRC-1, README.md, doctrine/design. No claim counterpart after comparison with RT-CLM-transaction-boundary and RT-CLM-derived-effects, which assess transaction/effect boundaries rather than advertised use.

• Agent context and memory so your AI tools have somewhere typed and persistent to share state among themselves and people. Structured, fully transparent and easy to maintain. --- README.md

EPI-CLM-period-learnings

Label: Evidence-grounded period learning lines

Claimed operation: propose concrete personal insights and external information grounded in period activity, merge existing lines, obtain corrections and write final learnings. Source-native learning-field claim, not by itself a claim of improved system capacity. SRC-1, examples/journal/skills/period-summary/SKILL.md, doctrine/design. No counterpart in supplied runtime Claims; distinct from the storage/use claim above.

Propose learnings based on actual period activity (children logs, achievements, events, and summaries), then merge with any learning lines already captured on the target period entry. Keep it concrete. Learnings are composed of personal insights and information from external sources like articles. --- examples/journal/skills/period-summary/SKILL.md