stacktrace-cli 0.11.0__tar.gz → 0.12.0__tar.gz

{stacktrace_cli-0.11.0 → stacktrace_cli-0.12.0}/CLAUDE.md

@@ -264,12 +264,55 @@ silently editing an accepted ADR breaks that contract.
   the user explicitly says "merge locally," "keep as branch," or
   "discard."
 
+## Writing documents
+
+Write all Markdown documents in ASD-STE100 Simplified Technical English.
+This includes specs, ADRs, READMEs, plans and release notes. The rule
+applies to every author, human or agent.
+
+- Write one idea in each sentence.
+- Use a maximum of 20 words in an instruction and 25 words in a
+  description.
+- Use a maximum of six sentences in a paragraph.
+- Use the active voice and the present tense.
+- Start an instruction with a verb, for example "Run the tests."
+- Use "a", "an" and "the". Do not remove them to make the text shorter.
+- Use one term for one thing, and use it the same way in all of the
+  document.
+- Use simple, common words. Do not use filler, such as "in order to" or
+  "utilize".
+- Do not use em-dash asides. Write a new sentence.
+- Do not change technical names, commands or code. Code blocks, tables and
+  quoted text do not have to obey these rules.
+
+Apply these rules to the text that you write or change. Do not rewrite an
+existing document only to change its style. Do not edit an accepted ADR to
+change its style.
+
+These rules are for authors. They are not a code review topic. Reviewers
+do not report style findings.
+
 ## Code Review Rules
 
+**The goal of review is to converge.** The reviewer and the fixer both work
+toward a PR that can merge in few rounds. Each round costs a full CI run and
+a review, and a finding on text that was already reviewed starts another
+round. Converging does not mean letting a serious defect through. The first
+review catches the glaring issues, and later reviews check only what changed.
+
 ### Reviewer
 
-- Review the full PR against its base branch. Report all qualifying
-  findings together; do not deliberately reserve findings for later rounds.
+- **The first review of a PR is the complete one.** Review the full PR
+  against its base branch. Report every qualifying finding together. Do not
+  reserve findings for later rounds.
+- **A later review covers only what changed since the previous reviewed
+  head.** That is the PR's new commits, plus existing code whose behavior
+  those commits change, such as a caller or dependency of changed code.
+  Changes that a merge or rebase brings in from the base branch are not part
+  of it.
+- **Do not report a defect that existed at the previous reviewed head**,
+  unless the new commits change its behavior. Exception: report a P0 or P1
+  even if it existed before, and label it a late discovery.
 - Report concrete, actionable defects with supported failure scenarios.
   Respect explicit scope decisions and accepted tradeoffs. Do not present
   speculative hardening or optional improvements as correctness defects.
@@ -280,12 +323,16 @@ silently editing an accepted ADR breaks that contract.
   - P3: low-priority suggestion.
 - Report only P0 and P1 findings. Do not post P2 or P3 findings, inline or
   in the review summary. Do not inflate priority to get a finding reported.
-- On subsequent reviews, verify earlier fixes and inspect their effects on
-  callers and dependencies. Older code within the PR remains reviewable.
+- A missing detail or an optional improvement is never P1. This applies to
+  code and to documentation. A P1 is something wrong: a defect, or a false
+  statement.
+- On later reviews, verify that the earlier fixes work. Check their effects on
+  callers and dependencies.
 - When review history supports it, identify a finding as:
   - Regression: introduced since the previous reviewed head.
-  - Late discovery: present at a previously reviewed head but not reported.
   - Unresolved: a previously reported defect remains.
+  - Late discovery: present at a previously reviewed head but not reported.
+    Only a P0 or P1 may be a late discovery.
   If the history is unavailable or ambiguous, say so rather than guessing.
 - Deduplicate by underlying defect and remedy, not by title. Refer to an
   existing thread for an unresolved defect instead of opening another one.
@@ -301,6 +348,13 @@ silently editing an accepted ADR breaks that contract.
   If its reasoning or priority is wrong, explain why rather than applying
   it solely because a reviewer requested it. Surface unresolved disputes
   for human judgment.
+- Decline a late discovery that is not P0 or P1. If the defect existed at the
+  previous reviewed head and the new commits do not change its behavior,
+  reply that the review rules exclude it, and do not fix it. A declined
+  finding does not count as a fix round.
+- Keep each fix to what its finding needs. Do not rewrite or extend nearby
+  code or text. A change beyond the finding gives the next review new
+  material to report, and the PR does not converge.
 - Fix the underlying invariant across relevant call sites. Test the
   failure class rather than only the reported example.
 - Preserve accepted fixes and regression tests when reviewing a rebased PR or

{stacktrace_cli-0.11.0 → stacktrace_cli-0.12.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: stacktrace-cli
-Version: 0.11.0
+Version: 0.12.0
 Summary: CLI for Stacktrace — Detection and Response platform for AI Agents.
 Project-URL: Homepage, https://stacktrace.ai
 Author-email: "Stacktrace AI, Inc" <founders@stacktrace.ai>
@@ -21,7 +21,7 @@ Requires-Python: >=3.11
 Requires-Dist: click>=8.1
 Requires-Dist: httpx<1.0.dev0,>=0.28.1
 Requires-Dist: openaca==0.10.1
-Requires-Dist: openaidr==0.6.3
+Requires-Dist: openaidr==0.7.0
 Requires-Dist: posthog<8,>=7.0
 Requires-Dist: pydantic<3,>=2.10
 Description-Content-Type: text/markdown
@@ -52,7 +52,7 @@ Requires Python 3.11 or newer.
 
 ```console
 $ stacktrace --version
-stacktrace 0.11.0 (openaca 0.10.1)
+stacktrace 0.12.0 (openaca 0.10.1)
 ```
 
 ## Python package namespace
@@ -250,6 +250,9 @@ Beta, and under active development.
 Devin sessions alike (Devin CLI and Devin Desktop's local agent). Session collection is
 [OpenAIDR](https://github.com/open-agent-security/openaidr)'s, and further agent
 kinds arrive with it. The daemon observes every kind it reads, sub-agents included.
+[Agent support](docs/agent-support.md) lists what works for each agent, rule by
+rule, and what cannot. Codex is next
+([spec](docs/specs/codex-agent-kind.md)).
 
 ## Built on
 

{stacktrace_cli-0.11.0 → stacktrace_cli-0.12.0}/README.md

@@ -24,7 +24,7 @@ Requires Python 3.11 or newer.
 
 ```console
 $ stacktrace --version
-stacktrace 0.11.0 (openaca 0.10.1)
+stacktrace 0.12.0 (openaca 0.10.1)
 ```
 
 ## Python package namespace
@@ -222,6 +222,9 @@ Beta, and under active development.
 Devin sessions alike (Devin CLI and Devin Desktop's local agent). Session collection is
 [OpenAIDR](https://github.com/open-agent-security/openaidr)'s, and further agent
 kinds arrive with it. The daemon observes every kind it reads, sub-agents included.
+[Agent support](docs/agent-support.md) lists what works for each agent, rule by
+rule, and what cannot. Codex is next
+([spec](docs/specs/codex-agent-kind.md)).
 
 ## Built on
 

{stacktrace_cli-0.11.0 → stacktrace_cli-0.12.0}/docs/adrs/0093-environment-cost-baselines-count-tokens-keyed-by-model.md

@@ -6,7 +6,7 @@ date: 2026-10-05
 supersedes: null
 superseded-by: null
 amends: null
-amended-by: 0099, 0113, 0114
+amended-by: 0099, 0113, 0114, 0129
 ---
 
 ## Context

{stacktrace_cli-0.11.0 → stacktrace_cli-0.12.0}/docs/adrs/0096-cost-usage-may-leave-as-token-totals-and-work-identifiers.md

@@ -6,7 +6,7 @@ date: 2026-10-06
 supersedes: null
 superseded-by: null
 amends: 0006
-amended-by: 0103, 0109, 0115, 0116
+amended-by: 0103, 0109, 0115, 0116, 0130
 ---
 
 ## Context

{stacktrace_cli-0.11.0 → stacktrace_cli-0.12.0}/docs/adrs/0099-cost-is-measured-in-tokens-and-relative-increases.md

@@ -6,7 +6,7 @@ date: 2026-10-06
 supersedes: 0094
 superseded-by: null
 amends: 0093
-amended-by: null
+amended-by: 0127
 ---
 
 ## Context

{stacktrace_cli-0.11.0 → stacktrace_cli-0.12.0}/docs/adrs/0104-usage-is-its-own-upload.md

@@ -6,7 +6,7 @@ date: 2026-10-08
 supersedes: null
 superseded-by: null
 amends: 0061
-amended-by: null
+amended-by: 0128
 ---
 
 ## Context

stacktrace_cli-0.12.0/docs/adrs/0127-the-monitor-shows-weighted-tokens-beside-raw-tokens.md

@@ -0,0 +1,121 @@
+---
+id: 0127
+title: The monitor shows weighted tokens beside raw tokens
+status: proposed
+date: 2026-10-15
+supersedes: null
+superseded-by: null
+amends: 0099
+amended-by: null
+---
+
+## Context
+
+The monitor's Usage tab adds fresh input, cache reads, cache writes and output
+one to one. On the measured machine 98% of those tokens were cache reads
+(ADR-0099). A cache read is billed at a tenth of a fresh input token on most
+Claude models, and at less on the newest. So the raw total mostly says how
+often a context was re-read. A long session that re-reads a large cache looks
+heavy. A session that sends the same context uncached on every call looks
+light.
+
+Fleet now judges and shows cost in dollars (Fleet ADR-0087). A developer who
+opens the monitor sees a raw count that moves differently from Fleet's figure.
+Nothing on the page says why.
+
+The product owner asked the monitor to show each kind of token on its own, and
+what each kind weighs, beside the raw count.
+
+ADR-0099 rejected tokens weighted by kind: "that is a price table by another
+name." Its concern was a table kept current for every model, and a figure read
+against an invoice it never matches. That conclusion does not hold for this
+use:
+
+- The table holds ratios between the kinds of one model, as the provider's
+  pricing page states them. It holds no price. The ratios change rarely: every
+  Claude model shares the same cache write and output ratios, and only four
+  models have cheaper cache reads.
+- A model the table does not list still gets a figure, at Anthropic's usual
+  ratios, marked estimated. ADR-0099's "a new model works on day one" holds.
+- A weighted token is still a token: one fresh input token of the model that
+  served it. It names no money and makes no claim about any invoice.
+- Rules, thresholds, baselines, evidence and the usual stay in raw tokens.
+  Only what the monitor shows changes.
+
+## Decision
+
+1. **The monitor shows weighted tokens beside raw tokens.** The Usage tab's
+   head states the range's weighted total under its raw total. A card lists each
+   kind of token with its raw count, the ratio it was weighed at, its weighted
+   count and its share of the weighted total. Each row of the table by model and
+   by agent carries its weighted count. The Live tab's session line carries the
+   session's weighted count beside its tokens.
+2. **One weighted token is one fresh input token of the model that served it.**
+   Each kind is multiplied by that model's published ratio to its own fresh
+   input, model by model, and then summed. On every Anthropic model: fresh input
+   ×1, 5-minute cache write ×1.25, 1-hour cache write ×2, output ×5. Cache reads
+   ×0.1, except ×0.05 on Opus 5.5 and Sonnet 5.5 and ×0.025 on Fable 5.1 and
+   Mythos 5.1. A cache write recorded without its lifetime weighs as a 5-minute
+   write.
+3. **The table lives in the client, as ratios** (`cost/weights.py`). A model is
+   looked up by the name `cost.models` gives it, without its snapshot date, and
+   without the effort suffix Devin records (`claude-opus-5-5-medium`).
+4. **A figure weighed by defaults is marked estimated.** A `claude-*` model the
+   table does not list, any other model (Devin's `swe-1-6`) and a response with
+   no model weigh at Anthropic's usual ratios, cache reads at ×0.1. No token is
+   dropped. The page says "estimated" and names the models.
+5. **Three charges are not weighted:** fast mode, US-only inference, and Haiku
+   5.5's tier for prompts over 100K tokens. Each is billed higher, so each
+   leaves a weighted figure under the bill, never over. The page says so.
+6. **Nothing else changes unit.** The chart, the runs, the components, the
+   usual, Session cost and its reasons, and every figure that leaves the
+   machine stay in raw tokens.
+
+What this amends: ADR-0099's rejection of "tokens weighted by kind", for what
+the monitor shows. ADR-0099's decision stands. Every cost figure is tokens, and
+a weighted token is a token. There is still no price table: this table holds
+ratios.
+
+## Alternatives considered
+
+- **Dollars on the machine.** Rejected by ADR-0099 and by this repo's rule that
+  cost is tokens, never dollars. It needs prices kept current, and invites a
+  comparison with an invoice that subscriptions and discounts never match.
+- **One reference model for every model.** Weigh each model's tokens against a
+  single model's fresh input, so Opus and Haiku add up on one scale. Rejected:
+  the ratio between two models is a ratio of their prices, so this is dollars in
+  disguise, with a reference price to keep current.
+- **Weigh only in Fleet**, as cost-notifications §6 proposed. Rejected: the
+  product owner asked for it on the machine. Fleet now shows dollars (Fleet
+  ADR-0087), so a monitor with raw tokens alone disagrees with Fleet by a factor
+  that moves with the mix of kinds, and nothing explains it.
+- **Show weighted tokens instead of raw ones.** Rejected: the rules, the usual
+  and Session cost read raw tokens, and the page must show the figure they
+  judge.
+- **Fold cache writes without a lifetime into 5-minute writes.** Rejected: that
+  changes the raw 5-minute figure. They get their own row when there are any.
+
+## Consequences
+
+- A reader sees why a large raw total can weigh little: cache reads count a
+  tenth or less.
+- The table is one more thing to keep current. A changed ratio needs a client
+  release. Until it ships, a listed model's figure is off, and a new model's
+  figure says estimated.
+- The machine's weighted tokens and Fleet's dollars are different scales. Within
+  one model they move together. Across models they do not, because each
+  model's weighted token is its own fresh input token.
+- The monitor weighs 1-hour writes at ×2, because transcripts split the two
+  lifetimes. cost-notifications §6 had Fleet weigh both at ×1.25.
+- `tests/cost/test_weights.py` holds each listed model's cache-read ratio,
+  checks that no kind of token is dropped, and checks that every model with a
+  known context window has ratios of its own.
+
+## When to revisit
+
+- When Anthropic changes a ratio, or a new model's cache reads differ from
+  ×0.1.
+- When OpenAIDR reads an agent whose provider publishes different ratios, such
+  as Codex.
+- If Session cost or the usual should judge weighted tokens. That changes the
+  rules, and needs its own ADR.

stacktrace_cli-0.12.0/docs/adrs/0128-usage-files-each-response-once-where-it-was-first-filed.md

@@ -0,0 +1,118 @@
+---
+id: 0128
+title: Usage files each response once, where it was first filed
+status: proposed
+date: 2026-10-15
+supersedes: null
+superseded-by: null
+amends: 0104
+amended-by: 0130, 0132
+---
+
+## Context
+
+An audit of the cost pipeline on 2026-10-15 found two ways one response reached
+Fleet twice. Both were reproduced through the real reader.
+
+- **Branches.** Claude Code 2.1.300's `/branch` writes a new session file
+  holding the conversation so far. Each record is copied whole under the new
+  session id, with its message id, usage and time, plus a `forkedFrom` field.
+  `--fork-session` likely does the same. OpenAIDR keys a response by its session
+  (its ADR-0014), and its token-usage spec leaves cross-session duplicates to
+  the consumer. One branch counted 1.66 times the tokens billed and raised a
+  false budget breach.
+- **A response read while it is written.** Claude Code writes a response a
+  block at a time, each record repeating its usage. The reader dates a response
+  by its newest record. A pass between two records files it a minute earlier
+  than the next pass does. Buckets merge by max and never move (ADR-0104), so
+  each minute kept a full copy. On this machine's transcripts that came to
+  about 2.4% of all tokens. One test response reached Fleet three times.
+
+## Decision
+
+A response is dated by its first turn: the first record of it that makes a
+turn, whose time never moves once written. A record of thinking alone makes no
+turn, so a response still being written with no turn yet waits for one. A
+final response, or one quiet for 10 minutes, is dated by its own time. The
+pass offers a session holding a waiting response again until it is filed.
+Every read dates a finished transcript the same way, so a store rebuilt from it
+sends Fleet the same bucket keys.
+
+The store records, per response, the session that first carried it and the
+minute it was dated in (`usage_responses`). `put_usage` files each response
+there, in the same transaction as the bucket merge. A later read files it in
+the same place, at its newest counts. A read of the same response from another
+session adds nothing.
+
+A response that no session has filed goes to the session that began first,
+with equal starts going by session id. A branch holds a subset of its
+original's records plus its own later ones, so it never began before the
+original. The machine's own figures use the same rank without a store: the
+cost rules through the ledger, the baseline's history of finished runs, the
+backfill's run names and components, and the monitor (`cost.copies.distinct`).
+
+Rows go at 30 days and on a switch to another Fleet, as buckets do.
+
+A database a build before this one wrote already has coverage, in
+`usage_buckets`, with nothing in `usage_responses` -- the table is additive
+(ADR-0057) and opens empty. Left alone, a branch read before its original is
+read again under this build would find nothing filed for the responses it
+copied and claim them as new. The first open under this build resets every
+covered kind's coverage once (`db.py`'s `_reclaim_usage_responses`), so the
+next backfill reads that history again and files it properly before any
+branch can; the buckets already merged are kept; the reread only adds to
+them. Guarded by `usage_responses` being empty, so it runs exactly once per
+database, ending for good at the first response the reread files.
+
+## Alternatives considered
+
+- **Skip records that carry `forkedFrom`, in OpenAIDR.** This is the exact
+  signal. It is rejected for now because OpenAIDR does not expose the field,
+  and changing that needs an upstream release and a new pin. Collection is not
+  ours to parse (ADR-0002). When OpenAIDR exposes it, the rank becomes a
+  fallback rather than the rule.
+- **Date a response by its first record.** That time is stable, but OpenAIDR
+  exposes only the newest record's. The first turn is the nearest stable time
+  it exposes; waiting for it costs a few seconds of upload latency.
+- **Date a response by the minute it was first read.** It is stable on one
+  machine, but a store rebuilt from the finished transcript would date it
+  differently, and Fleet could not match the resend to what it holds.
+- **Move buckets: send the old minute lowered, or zero.** Fleet applies a
+  lower revision as a negative change, so this would work on the wire. It is
+  rejected because it gives up the max merge. With the max merge, the backfill
+  and the pass commute, and an older read landing late cannot lower a count.
+- **Deduplicate only against the ledger's 36 hours.** That misses a branch of a
+  conversation older than the ledger, which is the likely
+  `--resume … --fork-session` case. It also re-ranks after every restart.
+
+## Consequences
+
+- Each response is sent to Fleet once. Branches no longer inflate totals,
+  budgets or runs on the machine or in Fleet.
+- A branch's own run no longer includes the conversation it inherited. Its run
+  and Live figures show only its own work.
+- The store holds one row per response for 30 days. A heavy machine makes a
+  few thousand responses a day.
+- Filing holds one object per response until its transaction ends. The
+  backfill files its history a slice of sessions at a time, at least 5,000
+  responses each, in rank order, and names runs after the last slice. On a
+  measured month of 60,000 responses its peak was 3.2 MiB, below the
+  13.5 MiB of filing the month's buckets at once. Nothing stays in memory:
+  the filings live in the store, and the daemon keeps only the ids of
+  sessions waiting for a response's first turn.
+- A response that straddled a minute and is read again just after upgrading may
+  still be counted in its old minute once. That is the stored state before this
+  change, not a new error.
+- A store that is lost and rebuilt has no filings. It dates each response the
+  same way, so its minute buckets match what Fleet holds. Its rows older than
+  seven days roll into hours, which Fleet cannot match to the minutes it
+  holds; that resend still counts twice until the upload says which store sent
+  a row.
+- A response read before its first turn waits up to 10 minutes. A response
+  that never gets a turn is filed late, never lost.
+
+## When to revisit
+
+When OpenAIDR exposes `forkedFrom`, or the first record's time of a response.
+Also when another agent kind's reader turns out to repeat a response id across
+distinct responses: the key assumes an id is unique within its kind.

stacktrace_cli-0.12.0/docs/adrs/0129-an-unfinished-response-counts-the-output-it-shows.md

@@ -0,0 +1,130 @@
+---
+id: 0129
+title: An unfinished response counts at least the output it shows
+status: proposed
+date: 2026-10-16
+supersedes: null
+superseded-by: null
+amends: 0093
+amended-by: null
+---
+
+## Context
+
+Claude Code writes a response one content block per record. Each record repeats
+the response's id and usage. Only the last record carries a stop reason, and
+only it states the whole output. The records before it state a small part.
+OpenAIDR marks a response `final` when it read that last record. It calls output
+on any other response a lower bound (its ADR-0014).
+
+Sub-agent transcripts mostly never get that last record. ADR-0093 counts what
+the transcripts record, on both sides of every comparison, and accepted that a
+run with many sub-agents counts its output low. Measured on one machine's
+transcripts on 2026-10-16:
+
+- 874 sub-agent responses were not final. They recorded 9,907 output tokens.
+- Their text and tool-call arguments came to 177,024 tokens at four characters
+  a token, read from the raw files.
+- Read through OpenAIDR they came to 105,802. The reader abridges any tool
+  argument over 1,000 characters to its first and last 400 (adr-sensor's
+  `truncate_middle`). The same cut applied to the raw files gives 105,802
+  exactly, on the same 874 responses.
+- So the recorded output was 9% of what those responses visibly wrote.
+
+Output is the kind the monitor's weighted tokens count heaviest, at five times
+fresh input (ADR-0127). Every cost figure on the machine, and every usage
+bucket sent to Fleet, undercounted each run that used sub-agents.
+
+## Decision
+
+A response that is not final counts the greater of its recorded output and
+what its turns show, at four characters a token.
+
+What a response shows is read from the assistant turns that carry its id in its
+own session:
+
+- A turn with no tool calls shows its text.
+- A turn that issued tool calls shows each call's arguments as JSON, and not its
+  text. The reader stands the placeholder `[Assistant used tools]` in for the
+  text of a record that holds only calls, and the model wrote none of it.
+- JSON keeps a character outside ASCII as written, so it counts once.
+
+A final response counts as recorded. Every other count of every response is
+unchanged.
+
+`cost.output.floored` applies the rule wherever sessions enter cost:
+
+- the ledger that the cost rules and the live baseline summaries read
+  (`Ledger.update`, once per change);
+- the usage recorder (`UsageRecorder.record`);
+- the usage backfill, on each slice it files;
+- the baseline backfill (`Baseline.backfill`);
+- the monitor's deduplicated view, which its range figures and Live cards read
+  (`snapshot_from`).
+
+A response's turns are its own session's, so the floor gives the same count
+before or after `cost.copies.distinct`. The floor keeps one integer per
+response and no text. `tests/cost/test_unfinished_output_counts_what_it_shows.py`
+asks every entry point through the real reader, and fails when a new module
+reads sessions cold or removes copies without being asked.
+
+A figure built from a response that is not final stays a lower bound, and the
+monitor still writes "at least". The usage recorder still waits up to 10
+minutes for a response with no turn yet (ADR-0128). Such a response shows
+nothing, so it files as recorded.
+
+This amends ADR-0093's "Recorded tokens on both sides". Baselines and rules
+count what the transcripts record, except that a response that is not final
+counts at least the output it shows.
+
+## Alternatives considered
+
+- **Count only final responses.** Rejected: a sub-agent whose responses never
+  finish would count no output at all. That is further from the truth than the
+  recorded fragment.
+- **Read the sub-agent's totals from the parent's Task tool result.** Rejected:
+  Claude Code's result for a sub-agent carries `isAsync`, `status`, `agentId`,
+  `description`, `resolvedModel`, `prompt`, `outputFile`, `canReadOutputFile`
+  and `canContinueAgent`, and no token count (13 results measured). Reading it
+  would also parse an agent format here, which ADR-0002 forbids.
+- **Three characters a token.** Rejected: on 425 final responses, whose output
+  is exact, three characters a token came out above the recorded output once,
+  at 1.07×. Four came out above it on none. A count above the true output
+  would stay in Fleet, because buckets merge by the greater (ADR-0104).
+- **Count the placeholder as text.** Rejected: the reader wrote it. It came to
+  23,276 characters on the measured responses, about 5% of the floor.
+- **Floor output in `Tokens.of`, where every reader counts.** Rejected: a
+  response alone carries no turns. The floor needs its session.
+- **Floor in the seam, `sessions/access.py`.** Rejected: the seam passes
+  OpenAIDR's sessions on as they are. `stacktrace sessions` shows what the
+  transcript recorded.
+- **Parse the transcript again for the content blocks.** Rejected by ADR-0002.
+
+## Consequences
+
+- Sub-agent output counts. On the measured machine the 874 unfinished
+  responses count 105,410 output tokens instead of 9,907.
+- The count is still a lower bound. Thinking does not show, and the reader
+  abridges long arguments. Through the real reader, on 388 final responses,
+  the floor came to a median of 0.25 of the recorded output and at most 0.79.
+- A count above a response's true output would stay in Fleet under the max
+  merge. Nothing measured did. A response made of long runs of one repeated
+  character could.
+- The floor rises as a response's records land, and its bucket rises with it.
+  A final record that arrives later replaces the floor when it is larger, and
+  whichever read lands last, the bucket keeps the larger.
+- Stored baseline rows catch up at the next daemon start: the backfill
+  summarizes every finished run again and replaces a row whose summary changed.
+- Usage Fleet already holds rises only when a read files that response again.
+  A store first opened by ADR-0128's build reads its covered history again
+  once (`db.py`'s `_reclaim_usage_responses`). Shipped in one release with this
+  change, that read refiles the last 30 days with the floor.
+- The ledger and each backfill slice walk the turns of unfinished responses
+  once. On a synthetic month of 60,000 responses, half of them unfinished, the
+  backfill's peak went from 3.0 MiB to 3.1 MiB.
+
+## When to revisit
+
+When Claude Code writes the last record of a sub-agent's responses, or OpenAIDR
+exposes a sub-agent's own totals. Also when another kind's reader marks
+responses not final and its turns can hold text the model did not write.