feat(always-on): measure the injected prefix, don't census the disk

[pa-arth]

What

The always-on breakdown answered "what did you put in your context" by walking files under ~/.claude. The context is not assembled from disk — Claude Code assembles it and records it, itemised, as typed attachment rows ahead of the first assistant turn. cc-audit read none of them.

Measured over 501 local sessions, --json --since-days 30:

component	before	after
skill listing	967 tok (skillCount: 15)	5,683 tok (skillCount: 78)
hook output	no field	948 tok
auto-memory	folded into projectClaudeMdTokens	2,990 tok, its own field
fixed prefix	no field	49,623 tok — 82% of the floor, and not yours to cut

Why it's a class, not four bugs

A disk census has no subscriber to a Claude Code release. Built-in skills ship inside the binary and are never on disk; the agent listing and deferred-tool delta move when the tool updates while nothing in ~/.claude does. It can't be wrong once and then fixed — it goes stale every release, silently.

It was also not stale in one direction, which the plan assumed:

• settings.json > skillOverrides was never read, so six skills the user had switched off were sized and billed every turn.
• readdirSync().isDirectory() is false for a symlinked directory, so four more were invisible.

Net under-reporting, per-skill over-reporting. "Scale the floor up" would not have fixed it — the membership set was wrong, not just the magnitude. The listing's own names[] is the authority now, and skillCarry[].loaded says so with a reason.

The disk census is retained and demoted: it attributes a measured block to your files; it no longer decides how big the block is. Plugin and user-command listings appear inside the injected skill listing, so they're slices of it, not addends to it — the old formula would have added them twice the moment the listing became measured.

Also

• ~/.claude/commands/** is walked at all now. Six opsx:* commands load into every listing; nothing had ever counted them.
• mcpDeferred is read from deferred_tools_delta in the sessions as they ran, not inferred from ENABLE_TOOL_SEARCH in the auditing process — a different process answering a different question.
• Unknown is null + a named reason, never 0. A session that recorded no attachments doesn't vote a 0 into any median, and doesn't vote on deferral either. That second conflation was a bug I wrote and a test caught mid-implementation.
• A negative reconciliation remainder fails and is reported per session. Clamping is how this defect class hides: the number stays plausible while the model underneath is wrong.

What I verified did NOT move

Run against origin/main at the same instant (the first attempt showed spurious drift purely because the two runs were 15 minutes apart and the trailing 30-day edge moved):

Zero drift across all 13 shared fields, plus spend.totalUsd and fluency.score identical to the last digit. Notably standingContextTokens (60,457 — already correct as of #27; the 40,487 in the plan was a pre-#27 figure) and pluginSkillTokens (a true zero: the walk runs, and the one enabled plugin ships no skills/ dir).

Tests

Fixtures are recorded from real transcripts by scripts/record-injected-fixture.mjs — structure, field names, string lengths and usage numbers preserved, content replaced with same-length filler. Never hand-authored: a fixture written from the design doc would have passed against a broken parser, because the doc had both the kind list and the per-kind content field wrong.

Four deliberate mutations were each caught by the specific test written for them: clamping fixedPrefix, sizing the JSON envelope, routing known-zero kinds to the residual, and letting an attachment-less session vote 0 into the medians.

402/403 pass. The one failure is pricingDrift (OpenAI rates vs LiteLLM), which fails identically on origin/main.

Breaking

Aggregate schemaVersion 9 → 10. skillDescriptionTokens is removed, not kept beside skillListingTokens — a field that keeps its name while changing meaning lets a downstream comparison silently mix two quantities, which is the defect class this PR closes.

⚠️ The backend reader is not in this PR yet. Per the spec, the contract and its reader ship together or neither ships — I'm tracking down the /v1/public/solo/data consumers now and will land that before this merges.

Spec: ~/repos/openspec/changes/standing-config-measured-not-censused/ (see findings-task-0.md — task 0 refuted four of this change's own premises before any code moved).

🤖 Generated with Claude Code

[greptile-apps]

Greptile Summary

The PR replaces disk-censused always-on estimates with measurements from first-turn transcript attachments and introduces aggregate schema version 10.

• Parses and reconciles injected skill, hook, MCP, deferred-tool, agent, residual, and fixed-prefix components.
• Retains disk inspection for attribution while using transcript listings as the loading authority.
• Adds recorded transcript fixtures, nullable unknown-state reporting, and expanded local and aggregate reporting.

Confidence Score: 4/5

The resumed-transcript turn mismatch should be fixed before merging because it can silently inflate the fixed-prefix and standing-context measurements.

Attachment collection can stop at a replayed assistant row that is subsequently deduplicated, while aggregation reconciles those attachments against a later retained turn containing additional conversation history.

Files Needing Attention: src/adapters/claudeCode.ts and src/alwaysOn.ts

Important Files Changed

Filename	Overview
src/adapters/claudeCode.ts	Collects first-turn attachment and user-message measurements, but closes collection before resumed-message deduplication.
src/alwaysOn.ts	Reworks always-on accounting around measured components; resumed transcripts can pair those components with a later retained turn.
src/injectedPrefix.ts	Defines attachment-field sizing, component folding, residual handling, and negative-remainder reconciliation.
src/aggregate.ts	Introduces schema version 10 with nullable measured-prefix fields and removes the old censused skill field.
src/report.ts	Presents the expanded measured always-on breakdown and reconciliation information.
scripts/record-injected-fixture.mjs	Records structurally faithful, length-preserving redacted transcript fixtures.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  T[Transcript rows] --> P[Parse pre-turn attachments]
  P --> C[Injected component counts]
  T --> U[First retained main-chain usage]
  C --> R[Per-session reconciliation]
  U --> R
  R --> M[Median always-on metrics]
  M --> A[Schema v10 aggregate and report]

Loading

Comments Outside Diff (1)

1. src/alwaysOn.ts, line 428 (link)

   Resumed prefix uses wrong turn

   When a resumed transcript's first assistant message is deduplicated and a later unseen turn is retained, firstMain supplies that later turn's history-inflated usage for attachments collected before the skipped turn, causing inflated fixed-prefix and standing-context costs while masking reconciliation mismatches.

   

_{Reviews (1): Last reviewed commit: "feat(always-on): measure the injected pr..." | Re-trigger Greptile}