Measurement

This page holds the latest measurement of dreamcatcher against the standard. The release checklist replaces it before each major or minor release, and a phase plans its work from it. Earlier measurements are in the Git history. The first was the baseline of 2026-09-30.

Measured on 2026-10-11 at 2d3c972, with the v5.5.0 release preparation.

Results

Criterion	Measured	Status	Issue
S1 Modules are cohesive	0 modules without one nameable responsibility	met
S2 Functions do one thing	0 functions without one nameable job	met
S3 Exports are used	33 names without an importer, all permitted	met
S4 Each thing is done one way	2 facts derived more than once	short	#452, #455
S5 Nothing is suppressed	0 noqa, 0 type: ignore, 3 platform pragmas	met
C1 One vocabulary	2 disagreements	short	#509, #511
C2 Enduring documents are true	0 new contradictions	met
C3 Documentation by purpose	4 kinds of page, kept apart	met
C4 Formats and contracts are versioned	state format 5, contract 1, both with policy	met
K1 Structural rules checked by machine	the 4 named rules checked	met
K2 Suite is fast and speaks plainly	24 seconds on a laptop, 100% branch coverage	met
K3 Tracker is current	42 issues reviewed; 5 corrected, none closed	met
K4 Every change reviewed to standard	10 pull requests answered; 7 name no issue	short	#510
E1 Rules, not cases	9 ledgered, all external; 4 unledgered	short	#419, #457
E2 Invariants by construction	61 optional fields, 0 coupled sets	met
E3 Concept economy	21 concepts; none added	met
E4 Symmetry	14 parallel operations; 1 uneven creation pair	short	#456

Shortfalls

• S4. The terminal and web views still derive issue-row status and evidence separately. That work waits on #455, which first settles how both kinds of agent work observe an issue. The rule that a round without an ending was interrupted once its daemon has gone remains in recovery, status and outcome reporting (#452).
• C1. The agent-facing contract calls dream:scout a conversation skill that follows the contract, and the skills guide routes a conversation label to a skill. The ontology defines only the assignment skill, and says a conversation recipe asks the agent to answer the user (#509).
• C1. The home page's terminal illustration writes an issue as #41, where dreamcatcher status writes GH41, and shows columns that the status view does not print (#511).
• K4. #499 to #505 came from one review of the documentation sites. Each answers every definition-of-done question, but answers the first with "No issue or spec". The standard does not say whether a change may serve neither (#510).
• E1. Four unledgered conditionals remain. The conversation scheduler returns early when no routes are configured, the templates place retry apart from the other controls, and the assignment card compares its status with display text (#419). The web also parses blocked evidence to recover issue numbers (#457).
• E4. Assignment creation is a class method, while conversation creation is a module function. Giving both creators one shape waits for the record-first setup design (#456, blocked by #425).

Notes

• Since v5.4.0. Ten pull requests merged. #491 records which specification phases are complete. #497 makes the placeholder commit and push of an assignment setup skip the repository's hooks. #499 to #505 introduce dream in the documentation, give the site a home page, a dusk scheme and new type, and move the skills into a "Customise skills" guide. #507 gives the web nature theme a dusk look and a moon when the system is in dark mode. Python source changed only in git.py.
• S1, S2. The largest modules are github.py (703 lines), agent_rounds.py (658) and web/app.py (598). They own GitHub projections and operations, round supervision, and HTTP routes respectively. git.py grew from 210 to 219 lines and still owns the Git operations. The longest functions are unchanged and each does one job:

Function	Lines	Its one job
cli._build_cli_parser	190	It holds the complete command-line grammar.
read_status_report	97	It composes the instance's status report.
AssignmentCreator.create	87	It follows assignment creation as the ontology lists it.
AgentRound.__init__	87	It starts a round as one transaction.
DreamcatcherDaemon.run	76	It follows the daemon lifecycle in the architecture.
Scheduler.tick	55	It follows the scheduler tick in the architecture.

• S3. The 33 names without a direct importer are unchanged. Document models, callback protocols and domain interfaces account for them.
• C1. The new conversation skill wording arrived with #499, whose answer to question 4 said it introduced no new term. Every other new word in the documentation, such as route, recipe and assignment skill, is the ontology's.
• C2. The architecture's account of assignment setup still holds after #497, since setup still makes an empty commit and pushes the branch. The known window in which a second daemon can run after the lock path is deleted, before the first daemon checks it, remains accepted as an edge case. No enduring claim newly contradicts the code.
• C3, C4. The new "Customise skills" guide sits with the how-to pages, and the home page links only user documentation. This release changes neither the state format nor the agent contract.
• K1. The architecture tests and subprocess lint rule still check the four named boundaries. No other rule has failed review more than once.
• K2. All 1,198 default tests passed with full branch coverage in 23.66 seconds. All seven browser tests passed in 7.73 seconds. The new tests name behaviours a user sees, such as a placeholder commit that skips failing hooks and the moon crossing the dusk sky.
• K3. The triage read all 42 open issues against the ontology, and rewrote #391, #404, #492, #494 and #496 to its words. #498 carries no type label. No open issue's work is complete or obsolete.
• K4. All ten pull requests answer the seven definition-of-done questions. #491, #497 and #507 each close an issue.
• E1. The nature script's day and night ranges and the stylesheet's colour scheme query apply general rules. The four previously reported special cases remain outstanding.
• E2. The count of 61 is unchanged, since no model changed.
• E3, E4. The ontology is unchanged. The two creators' shape remains uneven, and the existing count of parallel operations stands.

Method

• S1, S2. tools/measure_source.py reports module line counts and the ten longest functions. Review names each module's responsibility, largest first, and each long function's job, longest first.
• S3. Parse every module under src. Collect each module-level class, function, assignment and type alias whose name has no leading underscore. Remove a name when another source module imports it directly. Classify what remains as a document model, a view model a template consumes, or an interface the architecture names.
• S4, C1. Run the uncoded-consistency-review skill over the source, the enduring documents, the user documentation and the output a user sees.
• C2. Read the ontology, the architecture and the standard against the code, including recorded failure cases in open issues.
• K2. Time the default run on a laptop.
• E1. Read every conditional that names a particular situation against the ledger.
• E2. Count the class-level annotated fields, in dataclasses and pydantic models under src, whose type admits None. Read them for coupled sets.
• E4. Read the parallel names and the standing asymmetries in specs/2026-09-30-engineering-standard/agent-work-shape.md against the namespace map.