Lens output — embedded memory / context

Run AAS-20260820-CCDW-01. Step 6, embedded memory/context lens only. Boundary: subsystem-only, as declared in packet §3. Evidence tier: doc-grounded, fixed. Sources cited: SRC-1 only (SRC-2 is roster-only/negative and supports no finding here).

Standing scope note. Nothing in this run can reach implemented or above. Every "claimed" below means the vendor documentation asserts it; no implementation, run, or experiment was inspected. Statuses used: absent, inapplicable, uninspected, claimed.

Definitions applied. Memory read-back = material accumulated or changed through use returns to a later invocation or action. Static shipped material and ordinary current-run state are retained state, not read-back. Activation = evidence that delivered material changed behavior, not merely that it entered context.

1. Retained operative-part inventory

The run script has three distinct lives with different consumers and authority paths; they are split into rows 1a/1b/1c. OBJ-4 and OBJ-5 are split from each other because form, consumer, and surfacing differ.

# Part Storage substrate Representational form Persistence Lineage Producer Consumer Invalidation / regeneration Promotion path Status Evidence anchor
1a OBJ-1 in-runtime script text CMP-1 process memory symbolic (JavaScript) run lifetime authored by CMP-3 from the user prompt main-session Claude CMP-1 executor (BAP-1) ends with the process; regenerated by re-authoring → 1b (automatic), → 1c (manual s) claimed SRC-1#how-it-runs
1b OBJ-1 archived run script (CMP-8) file under the session dir in ~/.claude/projects/ symbolic (JavaScript) on disk; no documented bound, expiry, or deletion path one file per run; siblings are prior runs of the same session dir RTE-S1, automatic on every run human and Claude on request (BAP-9); becomes BAP-1 if relaunched uninspected — no retention limit, garbage collection, or staleness check documented edit → relaunch (BAP-9); or s → 1c claimed for the write; uninspected for lifecycle SRC-1#how-it-runs
1c OBJ-7 saved workflow command (CMP-7) .claude/workflows/ (project, repo-shared) or ~/.claude/workflows/ (personal) symbolic (JavaScript) + a command name across sessions; project-scoped or user-scoped promoted from a specific run's 1a via RTE-S2 human, pressing s and naming/locating it future sessions' command resolver, on /<name> (BAP-2) uninspected — no versioning, revalidation, staleness, or deletion documented; name collision is resolved at read time (project beats personal), not deduplicated at write time already the strongest documented form; no further promotion documented claimed SRC-1#save, #args
2 OBJ-3 script variables / intermediate results CMP-1 process memory in-language values, may hold natural-language agent output run-local; not documented as persisted derived from OBJ-2 and returned OBJ-4 the script the script; feeds RTE-X1 and the final synthesis ends with the process none documented claimed for existence; uninspected for whether it survives the process SRC-1#how-it-runs, #when-to-use
3 OBJ-4 agent result (prompt + recent tool calls + result) CMP-9 runtime store natural language + tool-call trace session-scoped; explicitly lost on exit one record per dispatched agent sub-agent (CMP-4) the runtime, for resume (RTE-C6); a human, via RTE-X6 invalidated by leaving the session — "the next session starts the workflow fresh" none documented — results are never distilled into a retained artifact claimed SRC-1#how-it-runs, #resume, #watch-the-run
4 OBJ-5 phase record (agent count, token total, elapsed) CMP-6 surface numeric/structured telemetry not documented beyond the run's visibility in /workflows aggregated over OBJ-4 runtime human only uninspected none claimed SRC-1#watch-the-run
5 OBJ-8 permission consent record project-scoped settings, or user settings in Auto mode configuration entry persistent across sessions produced as a by-product of an approval decision at RTE-C2 human's answer at the launch gate (RTE-S3) CMP-5 launch gate on later launches (BAP-5) uninspected — no revocation, expiry, or re-prompt-on-script-change documented none claimed SRC-1#approve
6 OBJ-12 tool allowlist user settings configuration persistent pre-existing user configuration; doc advises manual pre-population user, out of band inherited by every CMP-4 agent (BAP-6) uninspected uninspected claimed for inheritance; uninspected for whether mid-run prompts (RTE-C9) write back into it SRC-1#approve, #limits
7 OBJ-13 ultracode effort setting session configuration session-scoped, resets on new session set by /effort ultracode user CMP-3 planning (BAP-8) expires at session end by design none claimed SRC-1#ultracode
8 OBJ-14 session model selection session configuration session user, /model every agent, unless the script routes a stage elsewhere session end none claimed SRC-1#cost
9 CMP-10 disable setting (disableWorkflows, env var, managed settings) ~/.claude/settings.json, env, managed settings configuration persistent; org-wide via managed settings operator/admin configuration, not produced by use user or administrator whole feature (BAP-7) changed only by re-configuration; env var read at startup none claimed SRC-1#turn-off
10 CMP-11 /deep-research bundled script shipped with Claude Code symbolic (JavaScript) + command name ships with the product vendor-authored vendor command resolver, like 1c changed by product updates — (already a command) claimed; static shipped material, so retained state by definition SRC-1#bundled
11 OBJ-6 final report delivered into CMP-12 context natural language current conversation synthesized from OBJ-3 the script / its synthesizing stage main-session Claude + user (BAP-4) uninspected — no persistence of the report to a store is documented none documented claimed SRC-1#run-a-bundled-workflow
12 OBJ-9 claim, OBJ-11 vote/cross-check result inside OBJ-3 natural-language proposition; disposition value run-local produced by RTE-E3/RTE-E4 /deep-research agents and script the filter RTE-E5 and the synthesis RTE-E6 ends with the run none — surviving claims reach OBJ-6 as prose, not as retained records claimed SRC-1#bundled
13 Cross-run / cross-session finding store — — — — — — — — absent — no store of results, findings, summaries, or learned facts across runs is documented anywhere in SRC-1; the doctrine positively states a new session starts a workflow fresh SRC-1#resume

Structural observation (stated, not upgraded). Every documented durable retention in this subsystem holds a procedure (OBJ-1/OBJ-7) or an authority record (OBJ-8). Nothing documented retains findings. The doc's own framing agrees — "what's repeatable: the orchestration itself" (CLM-7). This is a description of the documented surface, not a claim that no finding-retention exists in the implementation.

2. Write side vs read-back, separated

2a. Write paths

Write path Write agency Function Consolidation / dedup / evolution / synthesis / invalidation / decay / promotion
RTE-S1 — archive the run script to CMP-8 automatic, unconditional ("every run") acquisition only No consolidation, no dedup (each run writes its own file even if the script is identical to the previous run's), no synthesis, no invalidation, no decay, no documented retention bound. Evolution is manual and out-of-band: the user may open, diff against a previous run's script, and edit the file. Promotion: the edited file becomes control flow only when relaunched (BAP-9 → BAP-1).
RTE-S2 — save a run's script as a command in CMP-7 manual (human selects the run in /workflows, presses s, chooses project vs personal, names it) curation — this is the subsystem's only curation act Selection, naming, and scope are all human decisions. No documented quality check, test, or validation gates the save. Name collision is not deduplicated at write time; it is resolved at read time (project entry wins). No versioning, no supersession, no deletion path documented. Promotion: run artifact → persistent command with BAP-2 force.
RTE-S3 — write the consent record OBJ-8 both, mode-dependent — in Default/accept-edits it is an explicit manual choice ("Yes, and don't ask again for <name> in <path>"); in Auto mode any Yes records consent in user settings, i.e. persistence is an automatic by-product of an ordinary approval acquisition of an authority fact No revocation, expiry, or re-prompt-on-change documented. Notably, nothing documents whether a changed script under the same workflow name re-triggers the prompt; if it does not, the record's scope has drifted from what was actually approved. Uninspected.
RTE-S5 — track each agent's result into CMP-9 automatic acquisition of raw traces Whole results retained; no consolidation, distillation, or summarization into a durable artifact documented. Decay is coarse and total: the store dies with the session.
Index maintenance for CMP-7 apparently automatic ("appear in / autocomplete alongside the bundled ones") index maintenance No index file, refresh trigger, or reconciliation is documented. Uninspected whether the index is derived on demand or materialized.
OBJ-12 allowlist writes uninspected — RTE-C9 documents that non-allowlisted shell/web/MCP calls prompt mid-run; it does not document whether answering that prompt writes back into the allowlist. If it does, this is a fourth accumulation-through-use write path and is currently unregistered.

Raw traces vs distilled artifacts. The subsystem holds raw traces (OBJ-4, OBJ-5) and, separately, a symbolic artifact (OBJ-1). There is no documented distillation step from traces into retained artifacts. RTE-S2 is not distillation: it promotes an artifact that already existed in symbolic form, unchanged, and it copies the procedure, never the findings. The one documented content-transforming pipeline (RTE-E3→RTE-E5 →RTE-E6) terminates in OBJ-6 delivered to a context window, not in any store.

2b. Actual read-back paths

Three paths meet the definition — accumulated-or-changed-through-use material returning to a later invocation.

  • RB-1 — Saved workflow command. RTE-S2 write → CMP-7 → RTE-C7 resolution of /<name> in future sessions, with BAP-2 force. Material: OBJ-7, produced by a specific past run. Status: claimed. This is the run's strongest read-back finding.
  • RB-2 — Archived run script re-read. RTE-S1 write → CMP-8 → a later read of a previous run's file to diff, edit, and relaunch (BAP-9). Material: OBJ-1 from an earlier run. Status: claimed for the availability and the edit-and-relaunch path; uninspected for how a later session locates the file (no index, search, or listing is documented — only the current run's path is handed over, by RTE-X4).
  • RB-3 — Persistent consent record. RTE-S3 write → OBJ-8 in project or user settings → consulted by CMP-5 at a later launch, suppressing the prompt (BAP-5). Status: claimed. This is the only read-back for which the documentation itself asserts a behavior change on the later invocation.

2c. Candidate paths classified as retained-state-only (with reasons)

  • RTE-C6 / RTE-S5 / CMP-9 resume cache. Explicitly retained state, not read-back. Resume returns cached results for the same run, inside the same session; the doc states that exiting Claude Code makes the next session start the workflow fresh. Ordinary current-run state by the run's definition. Classification: inapplicable as a read-back path; cross-session equivalent is absent by doctrine.
  • OBJ-3 script variables. Current-run process state. Retained state.
  • RTE-X4 path delivery at run start. The path handed to Claude is this run's archived script — current-run state, delivered forward within the run. Retained state. Only the re-read of a previous run's file (RB-2) is read-back.
  • OBJ-6 final report into CMP-12. Current-run output arriving in the current session's context. Retained state, not read-back; no later-invocation return documented.
  • CMP-11 /deep-research. Static shipped material. Retained state by definition, regardless of the fact that it resolves through the same registry as RB-1.
  • OBJ-12 tool allowlist as documented. Pre-existing user configuration inherited into agents — delivery of static configuration, not material accumulated through use. This classification is conditional: it flips to a read-back path if mid-run prompts write back into the allowlist, which is uninspected.
  • OBJ-13 ultracode setting. Session-scoped and explicitly resets; it cannot reach a later session.
  • OBJ-14 model selection, CMP-10 disable settings. Operator/admin configuration set deliberately out of band, not accumulated through use. Persistent, but not memory. Included in the inventory because they are retained operative parts; excluded from the read-back list on the accumulation criterion.

3. Context-route annotations (RTE-X1..RTE-X6)

Annotations only; routes are not re-inventoried or renamed. Direction is stated from the receiving agent's perspective.

RTE-X1 — script → sub-agent prompt

  • Read-back direction: push. The receiving sub-agent does not request the material; it is composed for it and installed at spawn. No pull affordance for the agent is documented.
  • Selection signal: script-authored code operating over OBJ-3/OBJ-2. Explicitly not model judgement at dispatch time.
  • Targeting: per-agent, addressed at spawn; each agent's prompt is individually composed and individually inspectable (RTE-X6).
  • Selection scope and budget: uninspected. No prompt-size budget, truncation rule, or selection policy is documented. The documented caps (≤16 concurrent, ≤1000 total) bound agent count, not per-prompt context.
  • Delivery and consumption point: the sub-agent's initial context at spawn (BAP-3, directive force for that worker's task).
  • Behavioral-faithfulness test: none documented. A human can read the prompt after the fact; nothing tests that the agent's behavior reflected it.
  • Memory classification: delivers current-run material. Retained-state delivery, not memory read-back.

RTE-X2 — sub-agent result → script variables

  • Read-back direction: the receiver is the script, not a model. Push, as a return value at task completion.
  • Selection signal: completion of the agent's task; the whole result returns. No filtering, summarization, or budget documented at this boundary.
  • Targeting: into OBJ-3, the dispatching script's variable space.
  • Scope and budget: uninspected.
  • Delivery and consumption point: script variables. This route puts material into no model's context; it enters one only if a later RTE-X1 composition includes it, or via RTE-X3. The packet's own anti-conflation check applies: retaining is not selecting.
  • Behavioral-faithfulness test: none.

RTE-X3 — final answer → main session

  • Read-back direction: push from the receiving session's perspective — the report "lands in your session" on completion; the session did not ask for it at that moment.
  • Selection signal: the script's terminal synthesis. Mixed form: code decides that one answer returns; a model stage decides what it says. For CMP-11 specifically, the content passed through RTE-E3 cross-check, RTE-E4 vote, and RTE-E5 filter.
  • Targeting: the originating session, CMP-12.
  • Selection scope and budget: one report; the explicit design point is context economy (CLM-3 — "Claude's context holds only the final answer"). No token budget documented.
  • Delivery and consumption point: the main session's context window, with advisory force (BAP-4).
  • Behavioral-faithfulness test: none. RTE-E5/RTE-E6 are content checks upstream of delivery — they filter unsurvived claims and attach citations. They test the material, not the receiver. I explicitly decline to read them as a faithfulness test, and decline to read arrival in context as activation.

RTE-X4 — run-script path → main session

  • Read-back direction: two-stage, mixed. The pointer is pushed at run start; the content is pulled on demand ("Claude receives the path when the run starts, so you can ask for it"). For a previous run's script (RB-2), it is pure pull.
  • Selection signal: positional for the push (the path of this run). Human intent for the pull. For a previous run's file: human intent alone — no index, search, listing, or recency ordering over CMP-8 is documented.
  • Targeting: the current session's Claude.
  • Selection scope and budget: one path string on the push. This is a pointer-not-payload design: the script body costs no context until requested.
  • Delivery and consumption point: main-session context; the file is read through ordinary file access when asked for. Force is material for relaunch (BAP-9) — advisory until relaunched, at which point it becomes BAP-1.
  • Behavioral-faithfulness test: none. Diffing against a previous run's script is a human comparison affordance, not a system check, and it compares scripts, never script behavior.

RTE-X5 — args → script global

  • Read-back direction: push at invocation; the receiver is the script, not a model. undefined if omitted.
  • Selection signal: the user names the input; main-session Claude converts it to structured data. That conversion is an uninspected transformation step — nothing documents that the structured args preserves what the user meant. I decline to infer semantic preservation across it.
  • Targeting: the invoked saved workflow's global namespace.
  • Scope and budget: uninspected.
  • Delivery and consumption point: script global, read directly with array/object methods without parsing.
  • Behavioral-faithfulness test: none.
  • Memory classification: invocation-time input. Retained state, not read-back.

RTE-X6 — run internals → human

  • Read-back direction: pull for the detail (the human opens /workflows, selects, drills into a phase, then into an agent). A thin push exists alongside it: the one-line progress summary in the task panel.
  • Selection signal: human navigation. No ranking, filtering, or salience mechanism documented.
  • Targeting: per-phase, then per-agent.
  • Selection scope and budget: documented and partial — phase agent count, token total, elapsed time; per agent, its prompt, recent tool calls, and result. "Recent" is a documented truncation of the trace, so what the human can inspect is not the full trace. j/k scrolling implies further display-level bounding.
  • Delivery and consumption point: the TUI / Desktop pane, read by a human; feeds the human control actions of RTE-C5 and the curation decision of RTE-S2.
  • Behavioral-faithfulness test: none. This is an inspection surface; it produces no verdict and gates nothing automatically.
  • Note: post-turn capture into CMP-9 is write-side maintenance (RTE-S5), not a second read-back point on this route.

4. Four-way separation table

Four independent findings per read-back path, each with its own evidence.

Path CONTEXT PRESENCE DEPLOYED WIRING ACTIVATION CAUSAL EFFECT
RB-1 saved workflow command claimed — the saved file exists at a documented location and "runs as /<name> in future sessions"; appears in / autocomplete. Evidence: SRC-1#save. claimed — resolution is documented, including precedence ("If a project workflow and a personal workflow share a name, the project one runs") and the args binding. Evidence: SRC-1#save, #args; authority BAP-2. uninspected — no run of a saved command was observed, and no mechanism is documented that would verify the saved orchestration behaves as the original run did. Prevents: any conclusion that saving a run reproduces that run's behavior, and any conclusion that the registry entry is what changed a later session's execution. uninspected — no comparison, ablation, or counterfactual. Prevents: any claim that saved workflows improve repeatability, consistency, or outcome quality in fact.
RB-2 archived run script re-read / diff / edit / relaunch claimed — "Every run writes its script to a file"; the file can be opened, and the path reaches Claude. Evidence: SRC-1#how-it-runs. claimed for the relaunch path — an edited file can be relaunched, converting BAP-9 material into BAP-1 control flow. uninspected for discovery of older files: no index, listing, or search over CMP-8 is documented, so the wiring from "a past run exists on disk" to "a later session finds it" rests on the human. Evidence: SRC-1#how-it-runs. uninspected — nothing documents that a re-read or diffed script changed any later behavior; the diff affordance is offered without any documented consequence. Prevents: any conclusion that the archive functions as a working memory of past orchestrations rather than a passive dump. uninspected. Prevents: any claim that archiving improves iteration, debugging, or script quality.
RB-3 persistent consent record OBJ-8 claimed — consent is written to project-scoped settings, or to user settings in Auto mode. Evidence: SRC-1#approve. claimed — the permission-mode table wires the record into the later launch decision; force is permissive (BAP-5). Evidence: SRC-1#approve. claimed — uniquely among the three, the doctrine asserts the behavior change directly: "later launches start without prompting", "skip this prompt for this workflow in this project from now on". This is a documented behavior change, still doc-only; it does not reach implemented or observed. Prevents (still): any conclusion that the suppression was observed, or that it behaves as documented at the host's actual version. uninspected — no observation, no counterfactual. Prevents: any claim about what consent accumulation does to real approval behavior, including the plausible-but-untested concern that Auto-mode by-product writes broaden consent beyond what the user intended.
CMP-9 resume cache (candidate) inapplicable as a read-back path — same run, same session; ordinary current-run state. Doctrine positively closes the cross-session case: "the next session starts the workflow fresh". Evidence: SRC-1#resume. inapplicable inapplicable inapplicable

Gap summary as required. All four causal cells and two of three activation cells are uninspected. Together they prevent every conclusion of the form "this subsystem's memory works" — including that saved workflows reliably re-run the same orchestration, that the archive supports iterative improvement, or that any retention mechanism produces a better outcome than not retaining. The only defensible summary from this run is that three retention-to-later-invocation paths are documented, one of which additionally has a documented behavior change.

5. Authority

Authority is reported by BAP-* ID, with consumer, channel, force, and horizon kept distinct from the family label.

  • RB-1 → BAP-2. Consumer: future sessions' command resolver plus Claude/user. Channel: CMP-7 registry lookup on /<name>. Force: enforcing on invocation — the saved orchestration is the control flow once invoked. Horizon: persistent across sessions; project-scoped (repo-shared, so it reaches other people's sessions) or user-scoped. Note the horizon-and-blast-radius asymmetry between the two save locations; it is a scope fact, not a quality fact.
  • RB-2 → BAP-9, then BAP-1. Consumer: Claude/user on request, then the runtime. Channel: the archived file, reached by the path from RTE-X4. Force: advisory while read; enforcing once relaunched — the transition is the relaunch, not the read. Horizon: on disk under the session directory, bound undocumented.
  • RB-3 → BAP-5. Consumer: the launch approval gate. Channel: the consent record in project or user settings. Force: permissive — it removes a check; it authorizes nothing new about what the script may do. Horizon: per workflow+project (default/acceptEdits) or per user (Auto).
  • Delivery of OBJ-6 → BAP-4. Consumer: main-session Claude and the user. Channel: RTE-X3. Force: advisory. Horizon: session/turn. Advisory force is the ceiling here; arrival in context is not activation.
  • Agent-side authority → BAP-6. The inherited allowlist plus forced acceptEdits is permissive/enforcing at the tool boundary and is independent of every memory path above. No retained memory artifact in this subsystem grants or widens tool authority.
  • Feature-level prohibition → BAP-7, and the session-scoped directive default BAP-8. Neither is memory read-back; both bound what any memory path can do.

Where I decline to infer.

  • Curation is not endorsement. RTE-S2 records a human's decision to keep a script. No quality check, test, review, or acceptance step gates the save. I decline to infer that a saved workflow was validated, that it is correct, or that its presence in CMP-7 carries any warrant beyond "someone pressed s". BAP-2's enforcing force describes what happens on invocation, not whether it should.
  • Lineage is not semantic preservation. The 1a → 1b → 1c chain is a lineage of copies and edits. Nothing in it establishes that the archived or saved script means the same thing as the run it came from — and the doc explicitly permits human editing between archive and relaunch (BAP-9), which is a documented break in preservation. Likewise the RTE-X5 conversion of user words into structured args is a transformation with no preservation check.
  • Operational continuation is not warrant. That a saved command keeps resolving in later sessions establishes persistence, not that its orchestration remains appropriate to a codebase that has since changed. No staleness or revalidation mechanism is documented.
  • CLM-1 is a warrant claim and is not supported by this lens. The doc asserts that adversarial review or multi-angle weighing "gives a more trustworthy result than a single pass". Per the packet's own anti-conflation check, RTE-E7/RTE-E8 are stated capabilities of the pattern, not shipped behavior. Nothing in the memory/context surface supplies evidence for the trustworthiness claim, and I do not treat the /deep-research filter-and-cite pipeline (CLM-2) as evidence for it.
  • No consolidation or import label exists here to lean on. There is no documented consolidation, merge, summarization, or import step in this subsystem. Had one existed, it would still not have established semantic preservation.

6. Limitations

Each paired with the exact conclusion it prevents.

  1. Doc-only evidence tier (doc-grounded, fixed). No implementation, run trace, or experiment. Prevents: any status above claimed for every finding in this lens; specifically prevents concluding that any write path actually writes, that any read-back path actually returns material, or that CMP-9 behaves as described.
  2. Version conflict, preserved. SRC-1 is pinned at 2026-06-03 / v2.1.154–v2.1.160; SRC-2 observed a v2.1.237 host. Prevents: any conclusion that the retention and read-back surface described here is the surface a current user encounters — including whether CMP-8 gained retention limits, or whether consent semantics changed.
  3. No script text was ever inspected (/deep-research's source is excluded by §3, and no example script exists in SRC-1). Prevents: any conclusion about how RTE-X1 actually selects material from OBJ-3 into a prompt — selection policy, ordering, size discipline, and truncation all remain unknown. This is the single largest gap in the context-assembly half of this lens.
  4. The sub-agent primitive's internals are out of the bundle (/en/sub-agents not fetched). Prevents: any conclusion about what a sub-agent loads besides its spawn prompt — project instruction files, skills, prior state, or its own memory. Consequently the claim that RTE-X1 is the only context route into a worker cannot be made.
  5. CMP-8 lifecycle undocumented. No retention bound, deletion, rotation, or index. Prevents: concluding that the archive is a usable memory surface over time rather than an unbounded write-only log; and prevents estimating whether RB-2 degrades with volume.
  6. Consent-record invalidation undocumented. Nothing states whether a changed script under an existing workflow name re-prompts. Prevents: concluding that OBJ-8's scope still matches what the human approved, and therefore prevents any safety conclusion about the persistent-consent path.
  7. Allowlist write-back undocumented. RTE-C9 documents the prompt but not its persistence. Prevents: settling whether OBJ-12 is static configuration (as classified here, provisionally) or a fourth accumulation-through-use read-back path.
  8. Agent-restart semantics undocumented. r restarts a running agent; nothing states whether the restarted agent receives anything from its prior attempt. Prevents: any conclusion about within-run retry memory, and leaves a possible intra-run retained-state route unregistered.
  9. OBJ-3 process survival undocumented (already flagged in packet §7, Loop C). Prevents: concluding that intermediate results are strictly ephemeral, and therefore prevents any confident statement about the actual data-at-rest footprint of a run.
  10. SRC-2 is roster-only and negative. Prevents: it supporting any positive finding in this lens; it is cited nowhere above as support.

7. Proposed new records

Returned to the orchestrator for canonical registration. No IDs minted.

  • PROPOSED NEW RECORD: state route (candidate, uninspected) — possible write-back from a mid-run tool permission prompt into the user's tool allowlist (OBJ-12). RTE-C9 registers the prompt; no record covers whether answering it persists. Registering it as an explicitly uninspected candidate keeps the retained-state-vs-read-back classification of OBJ-12 visible as provisional. — evidence anchor SRC-1#approve, #limits.
  • PROPOSED NEW RECORD: operative object — the / autocomplete command index derived over CMP-7 plus bundled commands. It is the discovery surface for RB-1 and is distinct from the registry files themselves; its maintenance semantics are undocumented. — evidence anchor SRC-1#save ("appear in / autocomplete alongside the bundled ones"), #bundled.
  • PROPOSED NEW RECORD: claim — retention claim: "Every run writes its script to a file under your session's directory in ~/.claude/projects/." Stated unconditionally, with no documented bound, expiry, or deletion. Worth a CLM-* because it is the load-bearing premise of RB-2 and BAP-9. — evidence anchor SRC-1#how-it-runs.
  • PROPOSED NEW RECORD: operative object / transformation point — main-session Claude's conversion of user-named input into the structured args value on RTE-X5. It is a model-performed semantic transformation on the invocation path of every saved workflow, and it has no preservation check. — evidence anchor SRC-1#args ("Claude passes the list as structured data").
  • PROPOSED NEW RECORD: gap — no cross-session or cross-run store of workflow results is documented; the doctrine positively asserts a fresh start in a new session. This is a documented absence rather than an uninspected one and deserves a record so later runs do not re-derive it. — evidence anchor SRC-1#resume.
  • PROPOSED NEW RECORD: gap — agent-restart (r) state semantics: whether a restarted agent inherits anything from its prior attempt is undocumented, leaving a possible intra-run retention route unregistered. — evidence anchor SRC-1#watch-the-run.